v1.0

model_executor/layers/fused_moe/__init__.py

@@ -0,0 +1,106 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from contextlib import contextmanager
+from typing import Any
+
+from vllm.model_executor.layers.fused_moe.config import FusedMoEConfig
+from vllm.model_executor.layers.fused_moe.fused_moe_method_base import (
+    FusedMoEMethodBase,
+)
+from vllm.model_executor.layers.fused_moe.layer import (
+    FusedMoE,
+    FusedMoeWeightScaleSupported,
+)
+from vllm.model_executor.layers.fused_moe.modular_kernel import (
+    FusedMoEActivationFormat,
+    FusedMoEPermuteExpertsUnpermute,
+    FusedMoEPrepareAndFinalize,
+)
+from vllm.model_executor.layers.fused_moe.shared_fused_moe import SharedFusedMoE
+from vllm.model_executor.layers.fused_moe.utils import activation_without_mul
+from vllm.triton_utils import HAS_TRITON
+
+_config: dict[str, Any] | None = None
+
+
+@contextmanager
+def override_config(config):
+    global _config
+    old_config = _config
+    _config = config
+    yield
+    _config = old_config
+
+
+def get_config() -> dict[str, Any] | None:
+    return _config
+
+
+__all__ = [
+    "FusedMoE",
+    "FusedMoEConfig",
+    "FusedMoEMethodBase",
+    "FusedMoeWeightScaleSupported",
+    "FusedMoEPermuteExpertsUnpermute",
+    "FusedMoEActivationFormat",
+    "FusedMoEPrepareAndFinalize",
+    "SharedFusedMoE",
+    "activation_without_mul",
+    "override_config",
+    "get_config",
+]
+
+if HAS_TRITON:
+    # import to register the custom ops
+    from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
+        BatchedDeepGemmExperts,
+    )
+    from vllm.model_executor.layers.fused_moe.batched_triton_or_deep_gemm_moe import (  # noqa: E501
+        BatchedTritonOrDeepGemmExperts,
+    )
+    from vllm.model_executor.layers.fused_moe.cutlass_moe import (
+        CutlassBatchedExpertsFp8,
+        CutlassExpertsFp8,
+        cutlass_moe_fp4,
+        cutlass_moe_fp8,
+    )
+    from vllm.model_executor.layers.fused_moe.deep_gemm_moe import DeepGemmExperts
+    from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
+        BatchedTritonExperts,
+    )
+    from vllm.model_executor.layers.fused_moe.fused_moe import (
+        TritonExperts,
+        fused_experts,
+        fused_topk,
+        get_config_file_name,
+        grouped_topk,
+    )
+    from vllm.model_executor.layers.fused_moe.triton_deep_gemm_moe import (
+        TritonOrDeepGemmExperts,
+    )
+
+    __all__ += [
+        "fused_topk",
+        "fused_experts",
+        "get_config_file_name",
+        "grouped_topk",
+        "cutlass_moe_fp8",
+        "cutlass_moe_fp4",
+        "CutlassExpertsFp8",
+        "CutlassBatchedExpertsFp8",
+        "TritonExperts",
+        "BatchedTritonExperts",
+        "DeepGemmExperts",
+        "BatchedDeepGemmExperts",
+        "TritonOrDeepGemmExperts",
+        "BatchedTritonOrDeepGemmExperts",
+    ]
+else:
+    # Some model classes directly use the custom ops. Add placeholders
+    # to avoid import errors.
+    def _raise_exception(method: str):
+        raise NotImplementedError(f"{method} is not implemented as lack of triton.")
+
+    fused_topk = lambda *args, **kwargs: _raise_exception("fused_topk")
+    fused_experts = lambda *args, **kwargs: _raise_exception("fused_experts")

model_executor/layers/fused_moe/all2all_utils.py

@@ -0,0 +1,160 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+
+import torch
+
+from vllm.distributed import (
+    get_ep_group,
+)
+from vllm.model_executor.layers.fused_moe.config import (
+    FusedMoEConfig,
+    FusedMoEParallelConfig,
+    FusedMoEQuantConfig,
+)
+from vllm.model_executor.layers.fused_moe.modular_kernel import (
+    FusedMoEPrepareAndFinalize,
+)
+from vllm.platforms import current_platform
+from vllm.utils.import_utils import has_deep_ep, has_pplx
+
+if current_platform.is_cuda_alike():
+    if has_pplx():
+        from .pplx_prepare_finalize import (
+            PplxPrepareAndFinalize,
+            pplx_hidden_dim_scale_bytes,
+        )
+    if has_deep_ep():
+        from .deepep_ht_prepare_finalize import DeepEPHTPrepareAndFinalize
+        from .deepep_ll_prepare_finalize import (
+            DEEPEP_QUANT_BLOCK_SHAPE,
+            DeepEPLLPrepareAndFinalize,
+        )
+
+
+def maybe_roundup_layer_hidden_size(
+    hidden_size: int,
+    act_dtype: torch.dtype,
+    moe_parallel_config: FusedMoEParallelConfig,
+) -> int:
+    """
+    Given layer hidden size and MoE configurations, round up hidden_size
+    if necessary.
+
+    Args:
+        hidden_size: Layer hidden-size
+        act_dtype: Data type of the layer activations.
+        moe_parallel_config: Fused MoE parallelization strategy configuration.
+
+    Return:
+        Rounded up hidden_size if rounding up is required based on the configs
+        and all2all backend.
+        Original hidden size otherwise.
+    """
+    if moe_parallel_config.use_deepep_ht_kernels:
+        hidden_size = DeepEPHTPrepareAndFinalize.maybe_roundup_layer_hidden_size(
+            hidden_size, act_dtype
+        )
+
+    if moe_parallel_config.use_deepep_ll_kernels:
+        hidden_size = DeepEPLLPrepareAndFinalize.maybe_roundup_layer_hidden_size(
+            hidden_size
+        )
+
+    return hidden_size
+
+
+def maybe_make_prepare_finalize(
+    moe: FusedMoEConfig,
+    quant_config: FusedMoEQuantConfig | None,
+) -> FusedMoEPrepareAndFinalize | None:
+    if not moe.moe_parallel_config.use_all2all_kernels:
+        return None
+
+    all2all_manager = get_ep_group().device_communicator.all2all_manager
+    assert all2all_manager is not None
+
+    prepare_finalize: FusedMoEPrepareAndFinalize | None = None
+
+    # TODO: could allow this now
+    assert not moe.use_flashinfer_cutlass_kernels, "Must be created in modelopt.py"
+
+    if moe.use_pplx_kernels:
+        assert quant_config is not None
+
+        hidden_dim_bytes, hidden_scale_bytes = pplx_hidden_dim_scale_bytes(
+            moe.max_num_tokens,
+            moe.hidden_dim,
+            moe.in_dtype,
+            quant_config.quant_dtype,
+            per_act_token_quant=quant_config.per_act_token_quant,
+            block_shape=quant_config.block_shape,
+        )
+
+        all_to_all_args = dict(
+            max_num_tokens=moe.max_num_tokens,
+            num_experts=moe.num_experts,
+            experts_per_token=moe.experts_per_token,  # topk
+            rank=all2all_manager.rank,
+            world_size=all2all_manager.world_size,
+            # dp_size actually means tp_size, bug in pplx kernels
+            dp_size=all2all_manager.tp_group.world_size,
+            hidden_dim=moe.hidden_dim,
+            hidden_dim_bytes=hidden_dim_bytes,
+            hidden_dim_scale_bytes=hidden_scale_bytes,
+        )
+
+        num_dispatchers = (
+            all2all_manager.world_size // all2all_manager.tp_group.world_size
+        )
+
+        # Intranode pplx a2a takes a group name while internode does not.
+        if not all2all_manager.internode:
+            all_to_all_args["group_name"] = all2all_manager.cpu_group.group_name
+
+        handle = all2all_manager.get_handle(all_to_all_args)
+
+        prepare_finalize = PplxPrepareAndFinalize(
+            handle,
+            max_num_tokens=moe.max_num_tokens,
+            num_local_experts=moe.num_local_experts,
+            num_dispatchers=num_dispatchers,
+        )
+    elif moe.use_deepep_ht_kernels:
+        assert moe.dp_size == all2all_manager.dp_world_size
+
+        all_to_all_args = dict()
+        handle = all2all_manager.get_handle(all_to_all_args)
+        prepare_finalize = DeepEPHTPrepareAndFinalize(
+            handle,
+            num_dispatchers=all2all_manager.world_size,
+            dp_size=all2all_manager.dp_world_size,
+            rank_expert_offset=all2all_manager.rank * moe.num_local_experts,
+        )
+
+    elif moe.use_deepep_ll_kernels:
+        assert quant_config is not None
+        all_to_all_args = dict(
+            max_num_tokens_per_dp_rank=moe.max_num_tokens,
+            token_hidden_size=moe.hidden_dim,
+            num_ep_ranks=all2all_manager.world_size,
+            num_global_experts=moe.num_experts,
+            num_local_experts=moe.num_experts // all2all_manager.world_size,
+        )
+        handle = all2all_manager.get_handle(all_to_all_args)
+
+        # Note: We may want to use FP8 dispatch just to reduce
+        # data movement.
+        use_fp8_dispatch = (
+            quant_config.quant_dtype == current_platform.fp8_dtype()
+            and quant_config.block_shape == DEEPEP_QUANT_BLOCK_SHAPE
+        )
+
+        prepare_finalize = DeepEPLLPrepareAndFinalize(
+            handle,
+            max_tokens_per_rank=moe.max_num_tokens,
+            num_dispatchers=all2all_manager.world_size,
+            use_fp8_dispatch=use_fp8_dispatch,
+        )
+
+    return prepare_finalize

model_executor/layers/fused_moe/batched_deep_gemm_moe.py

@@ -0,0 +1,406 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+
+import torch
+
+import vllm.model_executor.layers.fused_moe.modular_kernel as mk
+from vllm.forward_context import get_forward_context, is_forward_context_available
+from vllm.logger import init_logger
+from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
+from vllm.model_executor.layers.fused_moe.topk_weight_and_reduce import (
+    TopKWeightAndReduceDelegate,
+)
+from vllm.model_executor.layers.fused_moe.utils import _resize_cache
+from vllm.platforms import current_platform
+from vllm.triton_utils import tl, triton
+from vllm.utils.deep_gemm import (
+    DeepGemmQuantScaleFMT,
+    fp8_m_grouped_gemm_nt_masked,
+    get_mk_alignment_for_contiguous_layout,
+    is_deep_gemm_e8m0_used,
+)
+from vllm.utils.math_utils import cdiv, round_up
+
+logger = init_logger(__name__)
+
+
+def scales_shape_stride_dtype(
+    E: int, T: int, G: int, quant_scale_fmt: DeepGemmQuantScaleFMT
+) -> tuple[tuple[int, ...], tuple[int, ...], torch.dtype]:
+    shape = (E, T, G)
+    strides = (T * G, 1, T)
+    if quant_scale_fmt in [
+        DeepGemmQuantScaleFMT.FLOAT32,
+        DeepGemmQuantScaleFMT.FLOAT32_CEIL_UE8M0,
+    ]:
+        return shape, strides, torch.float32
+
+    assert quant_scale_fmt == DeepGemmQuantScaleFMT.UE8M0
+    shape = (E, T, cdiv(G, 4))
+    strides = (T * cdiv(G, 4), 1, T)
+    return shape, strides, torch.int32
+
+
+@triton.jit
+def _silu_mul_fp8_quant_deep_gemm(
+    # Pointers ------------------------------------------------------------
+    input_ptr,  # 16-bit activations (E, T, 2*H)
+    y_q_ptr,  # fp8 quantized activations (E, T, H)
+    y_s_ptr,  # 16-bit scales (E, T, G)
+    counts_ptr,  # int32 num tokens per expert (E)
+    # Sizes ---------------------------------------------------------------
+    H: tl.constexpr,  # hidden dimension (per output)
+    GROUP_SIZE: tl.constexpr,  # elements per group (usually 128)
+    # Strides for input (elements) ---------------------------------------
+    stride_i_e,
+    stride_i_t,
+    stride_i_h,
+    # Strides for y_q (elements) -----------------------------------------
+    stride_yq_e,
+    stride_yq_t,
+    stride_yq_h,
+    # Strides for y_s (elements) -----------------------------------------
+    stride_ys_e,
+    stride_ys_t,
+    stride_ys_g,
+    # Stride for counts (elements)
+    stride_counts_e,
+    # Numeric params ------------------------------------------------------
+    eps: tl.constexpr,
+    fp8_min: tl.constexpr,
+    fp8_max: tl.constexpr,
+    ceil_ue8m0: tl.constexpr,
+    # Meta ---------------------------------------------------------------
+    BLOCK: tl.constexpr,
+    NUM_STAGES: tl.constexpr,
+):
+    G = H // GROUP_SIZE
+
+    # map program id -> (e, g)
+    pid = tl.program_id(0)
+    e = pid // G
+    g = pid % G
+
+    e = e.to(tl.int64)
+    g = g.to(tl.int64)
+
+    # number of valid tokens for this expert
+    n_tokens = tl.load(counts_ptr + e * stride_counts_e).to(tl.int64)
+
+    cols = tl.arange(0, BLOCK).to(tl.int64)
+    mask = cols < BLOCK
+
+    base_input_offset = e * stride_i_e + g * GROUP_SIZE * stride_i_h
+    base_gate_offset = base_input_offset + cols * stride_i_h
+    base_up_offset = base_input_offset + H * stride_i_h + cols * stride_i_h
+    base_yq_offset = e * stride_yq_e + g * GROUP_SIZE * stride_yq_h + cols * stride_yq_h
+    base_ys_offset = e * stride_ys_e + g * stride_ys_g
+
+    for t in tl.range(0, n_tokens, num_stages=NUM_STAGES):
+        gate = tl.load(
+            input_ptr + base_gate_offset + t * stride_i_t, mask=mask, other=0.0
+        ).to(tl.float32)
+        up = tl.load(input_ptr + base_up_offset + t * stride_i_t, mask=mask, other=0.0)
+
+        gate = gate * (1.0 / (1.0 + tl.exp(-gate)))
+        y = gate * up
+
+        y_s = tl.maximum(tl.max(tl.abs(y)), eps) / fp8_max
+        if ceil_ue8m0:
+            y_s = tl.exp2(tl.ceil(tl.log2(y_s)))
+
+        y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
+
+        tl.store(y_q_ptr + base_yq_offset + t * stride_yq_t, y_q, mask=mask)
+        tl.store(y_s_ptr + base_ys_offset + t * stride_ys_t, y_s)
+
+
+def persistent_masked_m_silu_mul_quant(
+    y: torch.Tensor,  # (E, T, 2*H)
+    tokens_per_expert: torch.Tensor,  # (E,) number of valid tokens per expert
+    num_parallel_tokens=16,
+    group_size: int = 128,
+    quant_scale_fmt: DeepGemmQuantScaleFMT = DeepGemmQuantScaleFMT.FLOAT32,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Quantize silu(y[..., :H]) * y[..., H:] to FP8 with group per-token scales
+    y has shape (E, T, 2*H). The first half of the last dimension is
+    silu-activated, multiplied by the second half, then quantized into FP8.
+    We launch a fixed grid of threads to accommodate CUDA graphs. Let `P2`
+    be a parallelization factor for persistent_masked_m_silu_mul_quant over the
+    hidden dimension.
+
+    Let `expert_offsets = [0] + [num_tokens.cumsum()]` and
+    `total_tokens = expert_offsets[-1]`.
+    persistent_masked_m_silu_mul_quant launches `total_tokens x P2` number of
+    thread blocks. Each thread block contains `NUM_WARPS` warps.
+
+    Every thread block needs to find it's corresponding expert by warp-parallel scanning
+    over the `expert_offsets` array.
+
+    The i-th warp in the first thread block processes
+    `[i * warp_chunk_size, (i + 1) * warp_chunk_size]` groups
+    sequentially, where `warp_chunk_size = ((H / GROUP_SIZE) / P2) / NUM_WARPS`,
+    pipelining loads and computes.
+
+    The shared memory layout for 4 warps with a 2-stage pipeline for SiLU V2
+    can is visualized like so:
+
+                         stage0                              stage1
+    ┌─────┬───┬─────┬───┬─────┬───┬─────┬───┬─────┬───┬─────┬───┬─────┬───┬─────┬───┐
+    │gate0│up0│gate1│up1│gate2│up2│gate3│up3│gate0│up0│gate1│up1│gate2│up2│gate3│up3│
+    └─────┴───┴─────┴───┴─────┴───┴─────┴───┴─────┴───┴─────┴───┴─────┴───┴─────┴───┘
+
+    with the main difference between V1 and V2 being the global load
+    stride between warps, and between half-warps. Regarding the latter stride,
+    we assign the first half warp of every warp for `gate` loads and the second
+    half-warp to `up` loads.
+
+    Returns `(y_q, y_s)` where
+    * `y_q`: FP8 tensor, shape (E, T, H), same layout as y[..., :H]
+    * `y_s` depends on quant_scale_fmt,
+      - quant_scale_fmt == FLOAT32,
+         `y_s`: FP32 tensor, shape (E, T, H // group_size), strides (T*G, 1, T)
+      - quant_scale_fmt == E8M0,
+         `y_s`: Int32 tensor, shape (E, T, H // group_size // 4), strides (T*G, 1, T)
+      - quant_scale_fmt == E8M0_FLOAT32_SPARSE
+         `y_s`: FP32 tensor, shape (E, T, H // group_size), strides (T*G, 1, T)
+    Let NUM_WARPS be the number of warps in a single thread block and
+    `GROUP_SIZE = 128` be the size of the quantization group.
+    """
+    assert y.ndim == 3, "y must be (E, T, 2*H)"
+    E, T, H2 = y.shape
+    assert H2 % 2 == 0, "last dim of y must be even (2*H)"
+    H = H2 // 2
+    G = (H + group_size - 1) // group_size
+    assert H % 8 == 0, "H must be divisible by 8"
+    assert group_size == 128, "H must be divisible by 8"
+    assert tokens_per_expert.ndim == 1 and tokens_per_expert.shape[0] == E
+
+    tokens_per_expert = tokens_per_expert.to(device=y.device, dtype=torch.int32)
+
+    fp8_dtype = torch.float8_e4m3fn
+    y_q = torch.empty((E, T, H), dtype=fp8_dtype, device=y.device)
+
+    ys_shape, ys_strides, ys_dtype = scales_shape_stride_dtype(E, T, G, quant_scale_fmt)
+    y_s = torch.empty_strided(
+        ys_shape,
+        ys_strides,
+        dtype=ys_dtype,
+        device=y.device,
+    )
+
+    ceil_ue8m0 = quant_scale_fmt in [
+        DeepGemmQuantScaleFMT.FLOAT32_CEIL_UE8M0,
+        DeepGemmQuantScaleFMT.UE8M0,
+    ]
+
+    cuda_arch = current_platform.get_device_capability(
+        device_id=y.device.index
+    ).to_int()
+
+    if cuda_arch >= 80:
+        torch.ops._C.persistent_masked_m_silu_mul_quant(
+            y, tokens_per_expert, y_q, y_s, ceil_ue8m0
+        )
+    else:
+        stride_cnt_e = tokens_per_expert.stride()[0]
+
+        # Static grid over experts and H-groups.
+        # A loop inside the kernel handles the token dim
+        grid = (E * G,)
+        # strides (elements)
+        stride_i_e, stride_i_t, stride_i_h = y.stride()
+        stride_yq_e, stride_yq_t, stride_yq_h = y_q.stride()
+
+        f_info = torch.finfo(fp8_dtype)
+        fp8_max = f_info.max
+        fp8_min = f_info.min
+        eps: float = 1e-10
+        assert y_s.dtype == torch.float32, (
+            "_silu_mul_fp8_quant_deep_gemm does"
+            "not support {y_s.dtype} scales. Only torch.float32 supported."
+        )
+        _silu_mul_fp8_quant_deep_gemm[grid](
+            y,
+            y_q,
+            y_s,
+            tokens_per_expert,
+            H,
+            group_size,
+            stride_i_e,
+            stride_i_t,
+            stride_i_h,
+            stride_yq_e,
+            stride_yq_t,
+            stride_yq_h,
+            ys_strides[0],
+            ys_strides[1],
+            ys_strides[2],
+            stride_cnt_e,
+            eps,
+            fp8_min,
+            fp8_max,
+            ceil_ue8m0,
+            BLOCK=group_size,
+            NUM_STAGES=4,
+            num_warps=1,
+        )
+
+    return y_q, y_s
+
+
+class BatchedDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
+    def __init__(
+        self,
+        max_num_tokens: int,
+        num_dispatchers: int,
+        quant_config: FusedMoEQuantConfig,
+    ):
+        """
+        max_num_tokens: Maximum number of tokens from a DP Rank
+        num_dispatchers: The number of DP dispatchers.
+        quant_config: Quantization configuration
+        """
+        super().__init__(quant_config)
+        assert self.block_shape == get_mk_alignment_for_contiguous_layout()
+        assert self.quant_config.use_fp8_w8a8
+        self.max_num_tokens = max_num_tokens
+        self.num_dispatchers = num_dispatchers
+
+    @property
+    def activation_formats(
+        self,
+    ) -> tuple[mk.FusedMoEActivationFormat, mk.FusedMoEActivationFormat]:
+        return (
+            mk.FusedMoEActivationFormat.BatchedExperts,
+            mk.FusedMoEActivationFormat.BatchedExperts,
+        )
+
+    def supports_chunking(self) -> bool:
+        return False
+
+    def supports_expert_map(self) -> bool:
+        return False
+
+    def supports_packed_ue8m0_act_scales(self) -> bool:
+        """
+        DeepGemm supports packed ue8m0 activation scales format in devices == sm100
+        """
+        return is_deep_gemm_e8m0_used() and current_platform.is_device_capability(100)
+
+    def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
+        # Let PrepareAndFinalize::finalize() decide the impl.
+        return TopKWeightAndReduceDelegate()
+
+    def workspace_shapes(
+        self,
+        M: int,
+        N: int,
+        K: int,
+        topk: int,
+        global_num_experts: int,
+        local_num_experts: int,
+        expert_tokens_meta: mk.ExpertTokensMetadata | None,
+    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
+        # FIXME (varun): We should be able to dispatch only from the leader
+        # DP ranks in the case of TP > 1. At the moment, all the Ranks
+        # end up sending their tokens. This needs to be fixed.
+        num_dispatchers = self.num_dispatchers
+        num_experts = local_num_experts
+        max_num_tokens = M if self.max_num_tokens is None else self.max_num_tokens
+        workspace13 = (num_experts, max_num_tokens * num_dispatchers, max(K, N))
+        workspace2 = (num_experts, max_num_tokens * num_dispatchers, (N // 2))
+        output = (num_experts, max_num_tokens * num_dispatchers, K)
+        return (workspace13, workspace2, output)
+
+    def estimate_expected_m(
+        self, global_num_experts: int, max_tokens_per_expert: int, topk: int
+    ) -> int:
+        dp_meta = (
+            get_forward_context().dp_metadata
+            if is_forward_context_available()
+            else None
+        )
+        if dp_meta is None:
+            logger.warning_once(
+                "DPMetadata unavailable. Defaulting expected_m to "
+                f"{max_tokens_per_expert}.",
+                scope="local",
+            )
+            return max_tokens_per_expert
+
+        total_num_tokens = dp_meta.num_tokens_across_dp_cpu.sum().item()
+        total_num_tokens_replicated = total_num_tokens * topk
+
+        # Assume even load balancing
+        assert global_num_experts != 0
+        estimate = round_up(int(total_num_tokens_replicated // global_num_experts), 16)
+        # clamp estimate
+        estimate = max(estimate, 16)
+        estimate = min(max_tokens_per_expert, estimate)
+        return estimate
+
+    def apply(
+        self,
+        output: torch.Tensor,
+        hidden_states: torch.Tensor,
+        w1: torch.Tensor,
+        w2: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        activation: str,
+        global_num_experts: int,
+        expert_map: torch.Tensor | None,
+        a1q_scale: torch.Tensor | None,
+        a2_scale: torch.Tensor | None,
+        workspace13: torch.Tensor,
+        workspace2: torch.Tensor,
+        expert_tokens_meta: mk.ExpertTokensMetadata | None,
+        apply_router_weight_on_input: bool,
+    ):
+        assert expert_tokens_meta is not None
+        expert_num_tokens = expert_tokens_meta.expert_num_tokens
+
+        assert hidden_states.ndim == 3
+        assert self.block_shape is not None
+
+        a1q = hidden_states
+        _, N, K = w1.size()
+
+        assert w2.size(1) == K
+
+        E, max_num_tokens, N, K, _ = self.moe_problem_size(
+            hidden_states, w1, w2, topk_ids
+        )
+
+        workspace1 = _resize_cache(workspace13, (E, max_num_tokens, N))
+
+        expected_m = self.estimate_expected_m(
+            global_num_experts=global_num_experts,
+            max_tokens_per_expert=max_num_tokens,
+            topk=topk_ids.size(-1),
+        )
+
+        fp8_m_grouped_gemm_nt_masked(
+            (a1q, a1q_scale),
+            (w1, self.w1_scale),
+            workspace1,
+            expert_num_tokens,
+            expected_m,
+        )
+
+        quant_scale_fmt = DeepGemmQuantScaleFMT.from_oracle()
+        a2q, a2q_scale = persistent_masked_m_silu_mul_quant(
+            workspace1,
+            expert_num_tokens,
+            quant_scale_fmt=quant_scale_fmt,
+        )
+
+        fp8_m_grouped_gemm_nt_masked(
+            (a2q, a2q_scale),
+            (w2, self.w2_scale),
+            output,
+            expert_num_tokens,
+            expected_m,
+        )

model_executor/layers/fused_moe/batched_triton_or_deep_gemm_moe.py

@@ -0,0 +1,180 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+
+import vllm.model_executor.layers.fused_moe.modular_kernel as mk
+from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
+    BatchedDeepGemmExperts,
+)
+from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
+from vllm.model_executor.layers.fused_moe.fused_batched_moe import BatchedTritonExperts
+from vllm.utils.deep_gemm import get_mk_alignment_for_contiguous_layout
+
+
+class BatchedTritonOrDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
+    def __init__(
+        self,
+        max_num_tokens: int,
+        num_dispatchers: int,
+        quant_config: FusedMoEQuantConfig,
+        allow_deep_gemm: bool = False,
+    ):
+        super().__init__(quant_config)
+
+        self.batched_triton_experts = BatchedTritonExperts(
+            max_num_tokens=max_num_tokens,
+            num_dispatchers=num_dispatchers,
+            quant_config=self.quant_config,
+        )
+
+        self.allow_deep_gemm = (
+            allow_deep_gemm
+            and self.quant_config.use_fp8_w8a8
+            and self.block_shape == get_mk_alignment_for_contiguous_layout()
+        )
+
+        self.batched_deep_gemm_experts = (
+            BatchedDeepGemmExperts(
+                max_num_tokens=max_num_tokens,
+                num_dispatchers=num_dispatchers,
+                quant_config=self.quant_config,
+            )
+            if self.allow_deep_gemm
+            else None
+        )
+
+        assert (
+            self.batched_deep_gemm_experts is not None
+            or self.batched_triton_experts is not None
+        )
+
+    @property
+    def activation_formats(
+        self,
+    ) -> tuple[mk.FusedMoEActivationFormat, mk.FusedMoEActivationFormat]:
+        if self.batched_triton_experts is not None:
+            assert (
+                self.batched_deep_gemm_experts is None
+                or self.batched_deep_gemm_experts.activation_formats
+                == self.batched_triton_experts.activation_formats
+            )
+            return self.batched_triton_experts.activation_formats
+        else:
+            assert self.batched_deep_gemm_experts is not None
+            return self.batched_deep_gemm_experts.activation_formats
+
+    def supports_chunking(self) -> bool:
+        bdge = self.batched_deep_gemm_experts
+        bte = self.batched_triton_experts
+        return (bdge is None or bdge.supports_chunking()) and (
+            bte is None or bte.supports_chunking()
+        )
+
+    def supports_expert_map(self) -> bool:
+        bdge = self.batched_deep_gemm_experts
+        bte = self.batched_triton_experts
+        return (bdge is None or bdge.supports_expert_map()) and (
+            bte is None or bte.supports_expert_map()
+        )
+
+    def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
+        bdge = self.batched_deep_gemm_experts
+        bte = self.batched_triton_experts
+        bdge_war = bdge.finalize_weight_and_reduce_impl() if bdge else None
+        bte_war = bte.finalize_weight_and_reduce_impl() if bte else None
+        is_bdge_war = bdge_war is not None
+        is_bte_war = bte_war is not None
+
+        if is_bdge_war and is_bte_war:
+            assert bdge_war == bte_war, (
+                "Both implementations should agree on WeightAndReduce impls. "
+                f"Got bdge_war: {bdge_war}, and bte_war: {bte_war}"
+            )
+
+        if bdge_war is not None:
+            return bdge_war
+
+        assert bte_war is not None
+        return bte_war
+
+    def workspace_dtype(self, act_dtype: torch.dtype) -> torch.dtype:
+        return act_dtype
+
+    def workspace_shapes(
+        self,
+        M: int,
+        N: int,
+        K: int,
+        topk: int,
+        global_num_experts: int,
+        local_num_experts: int,
+        expert_tokens_metadata: mk.ExpertTokensMetadata | None,
+    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
+        # Note: the deep gemm workspaces are strictly larger than the triton
+        # workspaces so we can be pessimistic here and allocate for DeepGemm
+        # even if we fall back to triton later, e.g. if expert maps are set.
+        if self.allow_deep_gemm:
+            assert self.batched_deep_gemm_experts is not None
+            return self.batched_deep_gemm_experts.workspace_shapes(
+                M,
+                N,
+                K,
+                topk,
+                global_num_experts,
+                local_num_experts,
+                expert_tokens_metadata,
+            )
+        else:
+            assert self.batched_triton_experts is not None
+            return self.batched_triton_experts.workspace_shapes(
+                M,
+                N,
+                K,
+                topk,
+                global_num_experts,
+                local_num_experts,
+                expert_tokens_metadata,
+            )
+
+    def apply(
+        self,
+        output: torch.Tensor,
+        hidden_states: torch.Tensor,
+        w1: torch.Tensor,
+        w2: torch.Tensor,
+        topk_weights: torch.Tensor,
+        topk_ids: torch.Tensor,
+        activation: str,
+        global_num_experts: int,
+        expert_map: torch.Tensor | None,
+        a1q_scale: torch.Tensor | None,
+        a2_scale: torch.Tensor | None,
+        workspace13: torch.Tensor,
+        workspace2: torch.Tensor,
+        expert_tokens_meta: mk.ExpertTokensMetadata | None,
+        apply_router_weight_on_input: bool,
+    ):
+        experts = (
+            self.batched_deep_gemm_experts
+            if self.allow_deep_gemm
+            else self.batched_triton_experts
+        )
+        assert experts is not None
+        experts.apply(
+            output,
+            hidden_states,
+            w1,
+            w2,
+            topk_weights,
+            topk_ids,
+            activation,
+            global_num_experts,
+            expert_map,
+            a1q_scale,
+            a2_scale,
+            workspace13,
+            workspace2,
+            expert_tokens_meta,
+            apply_router_weight_on_input,
+        )

model_executor/layers/fused_moe/config.py

@@ -0,0 +1,916 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from dataclasses import dataclass
+from enum import IntEnum
+from typing import Optional, Union
+
+import torch
+
+import vllm.envs as envs
+from vllm.config import ParallelConfig
+from vllm.distributed import get_dp_group, get_tensor_model_parallel_rank
+from vllm.logger import init_logger
+from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import (
+    OCP_MX_DTYPES,
+    OCP_MX_Scheme,
+)
+from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
+from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
+from vllm.utils.import_utils import has_triton_kernels
+from vllm.utils.math_utils import cdiv
+
+logger = init_logger(__name__)
+
+if has_triton_kernels():
+    try:
+        from triton_kernels.matmul_ogs import PrecisionConfig
+    except ImportError:
+        logger.error(
+            "Failed to import Triton kernels. Please make sure your triton "
+            "version is compatible."
+        )
+
+
+def _get_config_dtype_str(
+    dtype: torch.dtype,
+    use_fp8_w8a8: bool = False,
+    use_int8_w8a16: bool = False,
+    use_int4_w4a16: bool = False,
+    ocp_mx_scheme: str | None = None,
+) -> str | None:
+    """
+    Return a string used to construct the filename that contains the
+    tuning info for a particular quantization scheme.  See
+    try_get_optimal_moe_config in fused_moe.py.
+    """
+    if use_fp8_w8a8:
+        return "fp8_w8a8"
+    elif use_int8_w8a16:
+        return "int8_w8a16"
+    elif use_int4_w4a16:
+        return "int4_w4a16"
+    elif ocp_mx_scheme is not None:
+        # The output of this function is passed to `try_get_optimal_moe_config`,
+        # and as we only simulate OCP MX execution in fused_moe for now,
+        # we will NOT look for `*,dtype=w_mxfp4_a_mxfp4.json` for now.
+        return None
+    elif dtype == torch.float:
+        # avoiding cases where kernel fails when float32 MoE
+        # use fp16/bfloat16 configs
+        return "float32"
+    return None
+
+
+def _quant_flags_to_group_shape(
+    quant_dtype: torch.dtype | str | None,
+    per_act_token_quant: bool,
+    per_out_ch_quant: bool,
+    block_shape: list[int] | None,
+) -> tuple[GroupShape | None, GroupShape | None]:
+    """
+    Convert MoE quantization flags into more generic GroupShapes.
+    """
+    a_shape: GroupShape | None
+    w_shape: GroupShape | None
+    if block_shape is not None:
+        assert not per_act_token_quant
+        assert not per_out_ch_quant
+        # TODO(bnell): this is not quite right for activations since first
+        # dim should be 1.
+        a_shape = GroupShape(row=block_shape[0], col=block_shape[1])
+        w_shape = GroupShape(row=block_shape[0], col=block_shape[1])
+    else:
+        w_shape = None
+        a_shape = None if quant_dtype is None else GroupShape.PER_TENSOR
+
+        if per_act_token_quant:
+            a_shape = GroupShape.PER_TOKEN
+
+        if per_out_ch_quant:
+            w_shape = GroupShape.PER_TOKEN
+
+    return a_shape, w_shape
+
+
+# The type of method in top-K routing
+# Please keep this in sync with the counterpart defined in https://github.com/flashinfer-ai/flashinfer/blob/main/include/flashinfer/trtllm/fused_moe/runner.h
+class RoutingMethodType(IntEnum):
+    # Default: Softmax -> TopK
+    Default = (0,)
+    # Renormalize: TopK -> Softmax
+    Renormalize = (1,)
+    # DeepSeekV3: Sigmoid -> RoutingBiasAdd -> Top2 in group -> Top4 groups
+    # -> Top8 experts from the Top4 groups
+    DeepSeekV3 = (2,)
+    # Llama4: Top1 -> Sigmoid
+    Llama4 = (3,)
+    # RenormalizeNaive: Softmax -> TopK -> Renormalize
+    RenormalizeNaive = (4,)
+    # TopK: TopK (no softmax)
+    TopK = (5,)
+    # Unspecified
+    Unspecified = 6.0
+
+
+@dataclass
+class FusedMoEQuantDesc:
+    """
+    A quantization descriptor for fused MoE ops. This class can describe
+    either activations or weights.
+    """
+
+    # The quantized type of this parameters.  None means unquantized or
+    # already quantized.
+    # TODO (bnell): use scalar_type instead of Union.
+    dtype: torch.dtype | str | None = None
+
+    # A field that describes the quantization group shape, from quant_utils.py.
+    #  * (-1, -1)   for per-tensor quantization
+    #  * (1, -1)    for per-row quantization
+    #  * (-1, 1)    for per-column quantization
+    #  * (128, 128) for 128x128 deepseek style block quantization
+    #  * (1, 128)   for deepseek style activation quantization
+    #               (i.e. per-token-per-group)
+    shape: GroupShape | None = None
+
+    # Quantization scales.
+    # TODO(bnell): maybe put PrecisionConfigs in subclass of QuantDesc?
+    scale: Union[torch.Tensor, "PrecisionConfig", None] = None
+
+    # Quantization alphas or gscales, used for nvfp4 types.
+    # TODO(bnell): put some of these in subclasses
+    alpha_or_gscale: torch.Tensor | None = None
+
+    # Zero points for int4/int8 types
+    zp: torch.Tensor | None = None
+
+    # Biases for GPT triton MoE
+    bias: torch.Tensor | None = None
+
+
+# TODO(bnell): have subclasses for specific moe methods?
+# e.g. for specific arguments bias, precision, etc.
+@dataclass
+class FusedMoEQuantConfig:
+    """
+    The FusedMoEQuantConfig contains all the quantization parameters for
+    a single FusedMoEMethodBase operation.  It consists of four
+    FusedMoEQuantDescs, one for each activation and set of weights.
+
+    Each FusedMoEMethodBase must implement a get_fused_moe_quant_config
+    method to construct a FusedMoEQuantConfig for use with that class.
+
+    FusedMoEQuant configs are only used for modular kernels, fused_experts
+    (from fused_moe.py), cutlass_moe_fp[48], rocm_aiter_fused_experts and
+    triton_kernel_moe_forward.  Other MoE methods can ignore the
+    FusedMoEQuantConfig (for now) and hardcode it to None.
+
+    There are currently some restrictions on what can be expressed:
+    - Most MoE ops only support similar quantization strategies for
+      each parameter, e.g. both weights must have the same GroupShape
+      and both activations must share the same GroupShape.  One exception to
+      this is the cutlass moe which allows per channel quantization on the
+      outputs.  Note: this restrictions are not always rigorously checked.
+    - Not all fused MoE functions support all the parameters, e.g. zero points,
+      global scales, alphas and biases are not universally supported.
+    - Fully general GroupShapes are not allowed.  Activations only support
+      per token, per tensor or K-blocked.
+    - Weights are not required to have a GroupShape since they have already
+      been quantized.
+
+    Other notes:
+    - PrecisionConfigs are specific to GPT OSS Triton.
+    - As a follow up it would probably make sense to subclass FusedMoEQuantDesc
+      or FusedMoEQuantConfig for particular FusedMoEMethodBase subclasses
+      so that only the required quantization parameters are used/stored.
+    """
+
+    # TODO(bnell) make sure a1_scales/a2_scales don't interfere with chunking
+    _a1: FusedMoEQuantDesc
+    _a2: FusedMoEQuantDesc
+    _w1: FusedMoEQuantDesc
+    _w2: FusedMoEQuantDesc
+
+    def __post_init__(self):
+        assert not self.per_act_token_quant or self.block_shape is None, (
+            "illegal quantization"
+        )
+
+    #
+    # Convenience accessors for various properties.
+    #
+
+    @property
+    def quant_dtype(self) -> torch.dtype | str | None:
+        return self._a1.dtype
+
+    @property
+    def is_quantized(self) -> bool:
+        return self.quant_dtype is not None
+
+    @property
+    def is_per_act_token(self) -> bool:
+        return self._a1.shape == GroupShape.PER_TOKEN
+
+    @property
+    def per_act_token_quant(self) -> bool:
+        return self._a1.shape == GroupShape.PER_TOKEN
+
+    @property
+    def per_out_ch_quant(self) -> bool:
+        return self._w1.shape == GroupShape.PER_TOKEN
+
+    @property
+    def is_per_tensor(self) -> bool:
+        return self._a1.shape == GroupShape.PER_TENSOR
+
+    @property
+    def block_shape(self) -> list[int] | None:
+        if (
+            self._a1.shape is not None
+            and self._a1.shape != GroupShape.PER_TENSOR
+            and self._a1.shape != GroupShape.PER_TOKEN
+        ):
+            return [self._a1.shape.row, self._a1.shape.col]
+        else:
+            return None
+
+    @property
+    def is_block_quantized(self) -> bool:
+        return self.block_shape is not None
+
+    @property
+    def a1_scale(self) -> torch.Tensor | None:
+        assert self._a1.scale is None or isinstance(self._a1.scale, torch.Tensor)
+        return self._a1.scale
+
+    @property
+    def a1_gscale(self) -> torch.Tensor | None:
+        return self._a1.alpha_or_gscale
+
+    @property
+    def a2_scale(self) -> torch.Tensor | None:
+        assert self._a2.scale is None or isinstance(self._a2.scale, torch.Tensor)
+        return self._a2.scale
+
+    @property
+    def a2_gscale(self) -> torch.Tensor | None:
+        return self._a2.alpha_or_gscale
+
+    @property
+    def w1_scale(self) -> torch.Tensor | None:
+        assert self._w1.scale is None or isinstance(self._w1.scale, torch.Tensor)
+        return self._w1.scale
+
+    @property
+    def w1_zp(self) -> torch.Tensor | None:
+        return self._w1.zp
+
+    @property
+    def w1_bias(self) -> torch.Tensor | None:
+        return self._w1.bias
+
+    @property
+    def w1_precision(self) -> Optional["PrecisionConfig"]:
+        assert self._w1.scale is None or isinstance(self._w1.scale, PrecisionConfig)
+        return self._w1.scale
+
+    @property
+    def g1_alphas(self) -> torch.Tensor | None:
+        return self._w1.alpha_or_gscale
+
+    @property
+    def w2_scale(self) -> torch.Tensor | None:
+        assert self._w2.scale is None or isinstance(self._w2.scale, torch.Tensor)
+        return self._w2.scale
+
+    @property
+    def w2_zp(self) -> torch.Tensor | None:
+        return self._w2.zp
+
+    @property
+    def w2_bias(self) -> torch.Tensor | None:
+        return self._w2.bias
+
+    @property
+    def w2_precision(self) -> Optional["PrecisionConfig"]:
+        assert self._w2.scale is None or isinstance(self._w2.scale, PrecisionConfig)
+        return self._w2.scale
+
+    @property
+    def g2_alphas(self) -> torch.Tensor | None:
+        return self._w2.alpha_or_gscale
+
+    @property
+    def use_fp8_w8a8(self) -> bool:
+        return self.quant_dtype == torch.float8_e4m3fn
+
+    @property
+    def use_int8_w8a8(self) -> bool:
+        return self.quant_dtype == torch.int8
+
+    @property
+    def use_int8_w8a16(self) -> bool:
+        return self._a1.dtype is None and self._w1.dtype == torch.int8
+
+    @property
+    def use_int4_w4a16(self) -> bool:
+        return self._a1.dtype is None and self._w1.dtype == "int4"
+
+    @property
+    def ocp_mx_scheme(self) -> str | None:
+        if not hasattr(self, "_ocp_mx_scheme"):
+            if (self._a1.dtype is not None and not isinstance(self._a1.dtype, str)) or (
+                self._w1.dtype is not None and not isinstance(self._w1.dtype, str)
+            ):
+                self._ocp_mx_scheme = None
+            else:
+                ocp_mx_scheme = OCP_MX_Scheme.from_quant_dtype(
+                    self._a1.dtype, self._w1.dtype
+                )
+
+                if ocp_mx_scheme is not None:
+                    ocp_mx_scheme = ocp_mx_scheme.value
+
+                self._ocp_mx_scheme = ocp_mx_scheme
+
+        return self._ocp_mx_scheme
+
+    @property
+    def use_mxfp4_w4a16(self) -> bool:
+        return self._a1.dtype is None and self._w1.dtype == "mxfp4"
+
+    @property
+    def use_nvfp4_w4a4(self) -> bool:
+        return self.quant_dtype == "nvfp4"
+
+    def config_name(self, dtype: torch.dtype) -> str | None:
+        """
+        Return a string used to construct the filename that contains the
+        tuning info for a particular quantization scheme.  See
+        try_get_optimal_moe_config in fused_moe.py.
+        """
+        return _get_config_dtype_str(
+            use_fp8_w8a8=self.use_fp8_w8a8,
+            use_int8_w8a16=self.use_int8_w8a16,
+            use_int4_w4a16=self.use_int4_w4a16,
+            ocp_mx_scheme=self.ocp_mx_scheme,
+            dtype=dtype,
+        )
+
+    def scale_shape(
+        self,
+        max_tokens: int,
+        hidden_dim: int,
+    ) -> tuple[int, int] | None:
+        """
+        Construct the proper activation scale shape for this
+        config.
+        """
+        if self.is_quantized:
+            if self.is_block_quantized:
+                assert self.block_shape is not None
+                _, block_k = self.block_shape
+                k_tiles = cdiv(hidden_dim, block_k)
+                return (max_tokens, k_tiles)
+            elif self.is_per_act_token:
+                return (max_tokens, 1)
+            else:
+                return (1, 1)
+        else:
+            return None
+
+    def batched_scale_shape(
+        self,
+        num_experts: int,
+        max_tokens: int,
+        hidden_dim: int,
+    ) -> tuple[int, int, int] | None:
+        """
+        Construct the proper activation batched scale shape for this
+        config, e.g. (num experts, *scale_shape).
+        """
+        if self.is_quantized:
+            scale_shape = self.scale_shape(max_tokens, hidden_dim)
+            assert scale_shape is not None
+            return (num_experts, *scale_shape)
+        else:
+            return None
+
+    @staticmethod
+    def make(
+        quant_dtype: torch.dtype | str | None = None,
+        per_act_token_quant: bool = False,
+        per_out_ch_quant: bool = False,
+        block_shape: list[int] | None = None,
+        w1_scale: Union[torch.Tensor, "PrecisionConfig", None] = None,
+        w2_scale: Union[torch.Tensor, "PrecisionConfig", None] = None,
+        a1_scale: torch.Tensor | None = None,
+        a2_scale: torch.Tensor | None = None,
+        g1_alphas: torch.Tensor | None = None,
+        g2_alphas: torch.Tensor | None = None,
+        a1_gscale: torch.Tensor | None = None,
+        a2_gscale: torch.Tensor | None = None,
+        w1_bias: torch.Tensor | None = None,
+        w2_bias: torch.Tensor | None = None,
+        w1_zp: torch.Tensor | None = None,
+        w2_zp: torch.Tensor | None = None,
+        weight_dtype: torch.dtype | str | None = None,
+    ) -> "FusedMoEQuantConfig":
+        """
+        General builder function for a FusedMoEQuantConfig.
+        - quant_dtype: Optional quantization type. None if activations are
+          unquantized or quantized prior to calling.  Note: "nvfp4", "mxfp4",
+          "mxfp6_e3m2", "mxfp6_e2m3" are the only valid string values
+          for quant_dtype.
+        - per_act_token_quant: Activations have per token quantization.
+        - per_out_ch_quant: Outputs have per channel quantization. (only
+          for cutlass).
+        - block_shape: Optional block size for block-wise quantization.
+          Incompatible with per_act_token and per_out_ch quant.
+        - w1_scale: Optional scale to be used for w1.
+        - w2_scale: Optional scale to be used for w2.
+        - a1_scale: Optional scale to be used for a1.
+        - a2_scale: Optional scale to be used for a2.
+        - g1_alphas: Optional global quantization scales for w1 (for nvfp4).
+        - g2_alphas: Optional global quantization scales for w2 (for nvfp4).
+        - a1_gscale: Optional global quantization scales for a1 (for nvfp4).
+        - a2_gscale: Optional global quantization scales for a2 (for nvfp4).
+        - w1_bias: Optional biases for w1 (GPT OSS Triton).
+        - w2_bias: Optional biases for w1 (GPT OSS Triton).
+        - w1_zp: Optional w1 zero points for int4/int8 quantization.
+        - w2_zp: Optional w2 zero points for int4/int8 quantization.
+        """
+        assert not isinstance(quant_dtype, str) or quant_dtype in {
+            "nvfp4",
+            "mxfp4",
+            "mxfp6_e3m2",
+            "mxfp6_e2m3",
+        }
+        assert not isinstance(weight_dtype, str) or weight_dtype in {
+            "nvfp4",
+            "mxfp4",
+            "mxfp6_e3m2",
+            "mxfp6_e2m3",
+        }
+
+        if weight_dtype is None:
+            weight_dtype = quant_dtype
+
+        a_shape, w_shape = _quant_flags_to_group_shape(
+            quant_dtype, per_act_token_quant, per_out_ch_quant, block_shape
+        )
+        quant_config = FusedMoEQuantConfig(
+            _a1=FusedMoEQuantDesc(quant_dtype, a_shape, a1_scale, a1_gscale),
+            _a2=FusedMoEQuantDesc(quant_dtype, a_shape, a2_scale, a2_gscale),
+            _w1=FusedMoEQuantDesc(
+                weight_dtype, w_shape, w1_scale, g1_alphas, w1_zp, w1_bias
+            ),
+            _w2=FusedMoEQuantDesc(
+                weight_dtype, w_shape, w2_scale, g2_alphas, w2_zp, w2_bias
+            ),
+        )
+        assert quant_config.per_act_token_quant == per_act_token_quant
+        assert quant_config.per_out_ch_quant == per_out_ch_quant
+        assert quant_config.block_shape == block_shape
+        return quant_config
+
+
+def fp8_w8a8_moe_quant_config(
+    w1_scale: torch.Tensor,
+    w2_scale: torch.Tensor,
+    a1_scale: torch.Tensor | None = None,
+    a2_scale: torch.Tensor | None = None,
+    per_act_token_quant: bool = False,
+    per_out_ch_quant: bool = False,
+    block_shape: list[int] | None = None,
+    a1_gscale: torch.Tensor | None = None,
+    a2_gscale: torch.Tensor | None = None,
+    g1_alphas: torch.Tensor | None = None,
+    g2_alphas: torch.Tensor | None = None,
+) -> FusedMoEQuantConfig:
+    """
+    Construct a quant config for fp8 activations and fp8 weights.
+    """
+    return FusedMoEQuantConfig.make(
+        torch.float8_e4m3fn,
+        w1_scale=w1_scale,
+        g1_alphas=g1_alphas,
+        w2_scale=w2_scale,
+        g2_alphas=g2_alphas,
+        a1_scale=a1_scale,
+        a1_gscale=a1_gscale,
+        a2_scale=a2_scale,
+        a2_gscale=a2_gscale,
+        per_act_token_quant=per_act_token_quant,
+        per_out_ch_quant=per_out_ch_quant,
+        block_shape=block_shape,
+    )
+
+
+def int8_w8a8_moe_quant_config(
+    w1_scale: torch.Tensor,
+    w2_scale: torch.Tensor,
+    a1_scale: torch.Tensor | None,
+    a2_scale: torch.Tensor | None,
+    per_act_token_quant: bool = False,
+) -> FusedMoEQuantConfig:
+    """
+    Construct a quant config for int8 activations and int8 weights.
+    """
+    return FusedMoEQuantConfig.make(
+        torch.int8,
+        w1_scale=w1_scale,
+        w2_scale=w2_scale,
+        a1_scale=a1_scale,
+        a2_scale=a2_scale,
+        per_act_token_quant=per_act_token_quant,
+        per_out_ch_quant=False,
+        block_shape=None,
+    )
+
+
+def mxfp4_w4a16_moe_quant_config(
+    w1_scale: Union[torch.Tensor, "PrecisionConfig"],
+    w2_scale: Union[torch.Tensor, "PrecisionConfig"],
+    w1_bias: torch.Tensor | None = None,
+    w2_bias: torch.Tensor | None = None,
+) -> FusedMoEQuantConfig:
+    """
+    Construct a quant config for unquantized activations and mxfp4 weights.
+    """
+    return FusedMoEQuantConfig(
+        _a1=FusedMoEQuantDesc(),
+        _a2=FusedMoEQuantDesc(),
+        _w1=FusedMoEQuantDesc("mxfp4", None, w1_scale, None, None, w1_bias),
+        _w2=FusedMoEQuantDesc("mxfp4", None, w2_scale, None, None, w2_bias),
+    )
+
+
+def mxfp4_mxfp8_moe_quant_config(
+    w1_scale: Union[torch.Tensor, "PrecisionConfig"],
+    w2_scale: Union[torch.Tensor, "PrecisionConfig"],
+    a1_scale: torch.Tensor | None = None,
+    a2_scale: torch.Tensor | None = None,
+    w1_bias: torch.Tensor | None = None,
+    w2_bias: torch.Tensor | None = None,
+    block_shape: list[int] | None = None,
+) -> FusedMoEQuantConfig:
+    """
+    Construct a quant config for mxfp4 activations and mxfp4 weights.
+    """
+    return FusedMoEQuantConfig(
+        _a1=FusedMoEQuantDesc("mxfp8"),
+        _a2=FusedMoEQuantDesc("mxfp8"),
+        _w1=FusedMoEQuantDesc("mxfp4", None, w1_scale, None, None, w1_bias),
+        _w2=FusedMoEQuantDesc("mxfp4", None, w2_scale, None, None, w2_bias),
+    )
+
+
+def ocp_mx_moe_quant_config(
+    quant_dtype: str,
+    w1_scale: Union[torch.Tensor, "PrecisionConfig"],
+    w2_scale: Union[torch.Tensor, "PrecisionConfig"],
+    weight_dtype: str | None = None,
+    a1_scale: torch.Tensor | None = None,
+    a2_scale: torch.Tensor | None = None,
+    w1_bias: torch.Tensor | None = None,
+    w2_bias: torch.Tensor | None = None,
+    block_shape: list[int] | None = None,
+) -> FusedMoEQuantConfig:
+    """
+    Construct a quant config for mxfp4 activations and mxfp4 weights.
+    """
+    assert quant_dtype in OCP_MX_DTYPES
+    return FusedMoEQuantConfig.make(
+        quant_dtype=quant_dtype,
+        weight_dtype=weight_dtype,
+        w1_scale=w1_scale,
+        w2_scale=w2_scale,
+        a1_scale=a1_scale,
+        a2_scale=a2_scale,
+        w1_bias=w1_bias,
+        w2_bias=w2_bias,
+        per_act_token_quant=False,
+        per_out_ch_quant=False,
+        block_shape=block_shape,
+    )
+
+
+def nvfp4_moe_quant_config(
+    g1_alphas: torch.Tensor,
+    g2_alphas: torch.Tensor,
+    a1_gscale: torch.Tensor,
+    a2_gscale: torch.Tensor,
+    w1_scale: torch.Tensor,
+    w2_scale: torch.Tensor,
+) -> FusedMoEQuantConfig:
+    """
+    Construct a quant config for mxfp4 activations and nvp4 weights.
+    """
+    return FusedMoEQuantConfig.make(
+        "nvfp4",
+        w1_scale=w1_scale,
+        w2_scale=w2_scale,
+        a1_gscale=a1_gscale,
+        a2_gscale=a2_gscale,
+        g1_alphas=g1_alphas,
+        g2_alphas=g2_alphas,
+        per_act_token_quant=False,
+        per_out_ch_quant=False,
+        block_shape=None,
+    )
+
+
+def int4_w4a16_moe_quant_config(
+    w1_scale: torch.Tensor,
+    w2_scale: torch.Tensor,
+    w1_zp: torch.Tensor | None,
+    w2_zp: torch.Tensor | None,
+    block_shape: list[int] | None = None,
+) -> FusedMoEQuantConfig:
+    """
+    Construct a quant config for 16-bit float activations and int4 weights.
+    Note: Activations are pre-quantized.
+    """
+    group_shape = GroupShape(*block_shape) if block_shape is not None else None
+    return FusedMoEQuantConfig(
+        _a1=FusedMoEQuantDesc(shape=group_shape),
+        _a2=FusedMoEQuantDesc(shape=group_shape),
+        _w1=FusedMoEQuantDesc("int4", group_shape, w1_scale, None, w1_zp),
+        _w2=FusedMoEQuantDesc("int4", group_shape, w2_scale, None, w2_zp),
+    )
+
+
+def int8_w8a16_moe_quant_config(
+    w1_scale: torch.Tensor,
+    w2_scale: torch.Tensor,
+    w1_zp: torch.Tensor | None,
+    w2_zp: torch.Tensor | None,
+    block_shape: list[int] | None = None,
+) -> FusedMoEQuantConfig:
+    """
+    Construct a quant config for 16-bit float activations and int8 weights.
+    Note: Activations are pre-quantized.
+    """
+    group_shape = GroupShape(*block_shape) if block_shape is not None else None
+    return FusedMoEQuantConfig(
+        _a1=FusedMoEQuantDesc(shape=group_shape),
+        _a2=FusedMoEQuantDesc(shape=group_shape),
+        _w1=FusedMoEQuantDesc(torch.int8, group_shape, w1_scale, None, w1_zp),
+        _w2=FusedMoEQuantDesc(torch.int8, group_shape, w2_scale, None, w2_zp),
+    )
+
+
+def biased_moe_quant_config(
+    w1_bias: torch.Tensor | None,
+    w2_bias: torch.Tensor | None,
+) -> FusedMoEQuantConfig:
+    """
+    Construct a quant config for unquantized activations with biases.
+    """
+    return FusedMoEQuantConfig(
+        _a1=FusedMoEQuantDesc(),
+        _a2=FusedMoEQuantDesc(),
+        _w1=FusedMoEQuantDesc(bias=w1_bias),
+        _w2=FusedMoEQuantDesc(bias=w2_bias),
+    )
+
+
+# A FusedMoEQuantConfig constant for an unquantized MoE op.
+FUSED_MOE_UNQUANTIZED_CONFIG: FusedMoEQuantConfig = FusedMoEQuantConfig.make()
+
+
+@dataclass
+class FusedMoEParallelConfig:
+    tp_size: int
+    dp_size: int
+    ep_size: int
+    tp_rank: int
+    dp_rank: int
+    ep_rank: int
+
+    use_ep: bool  # whether to use EP or not
+    all2all_backend: str  # all2all backend for MoE communication
+
+    @property
+    def use_all2all_kernels(self):
+        return self.dp_size > 1 and self.use_ep
+
+    @property
+    def use_pplx_kernels(self):
+        return self.use_all2all_kernels and self.all2all_backend == "pplx"
+
+    @property
+    def use_deepep_ht_kernels(self):
+        return (
+            self.use_all2all_kernels
+            and self.all2all_backend == "deepep_high_throughput"
+        )
+
+    @property
+    def use_deepep_ll_kernels(self):
+        return self.use_all2all_kernels and self.all2all_backend == "deepep_low_latency"
+
+    @staticmethod
+    def flatten_tp_across_dp(
+        tp_size: int, dp_size: int, dp_rank: int
+    ) -> tuple[int, int]:
+        tp_rank = 0 if tp_size == 1 else get_tensor_model_parallel_rank()
+        # There are actually dp_size * tp_size devices. Update tp_size
+        # and tp_rank so we shard across all devices.
+        flatten_tp_size = dp_size * tp_size
+        flatten_tp_rank = dp_rank * tp_size + tp_rank
+        return flatten_tp_size, flatten_tp_rank
+
+    @staticmethod
+    def make(
+        tp_size_: int, dp_size_: int, vllm_parallel_config: ParallelConfig
+    ) -> "FusedMoEParallelConfig":
+        """
+        Determine MoE parallel configuration. Based on the input `tp_size_`,
+        `dp_size_` and vllm's parallel config, determine what
+        level's of parallelism to use in the fused moe layer.
+
+        Args:
+            tp_size_ (int): `tp_size` passed into the FusedMoE constructor.
+            dp_size_ (int): `dp_size` passed into the FusedMoE constructor.
+            vllm_parallel_config (ParallelConfig): vLLM's parallel config
+                object which contains the `enable_expert_parallel` flag.
+
+        Examples:
+            When there is no parallelism requested,
+            i.e. `tp_size_` = `dp_size_` = 1, we simply return the sizes
+            unaltered and the ranks set to 0.
+
+            Expert Parallelism is considered only when either `dp_size_` or
+            `tp_size_` is non trivial.
+
+            When TP = 2, DP = 1 and EP = False, the configuration on different
+            devices:
+
+            - device 0 : TP = {2, 0} DP = {1, 0} EP = {1, 0} //
+                legend : {size, rank}
+            - device 1 : TP = {2, 1} DP = {1, 0} EP = {1, 0}
+            - Comment : Tensors are sharded across 2 devices.
+
+            When TP = 1, DP = 2 and EP = False, the configuration on different
+                devices:
+
+            - device 0 : TP = {2, 0} DP = {2, 0} EP = {1, 0}
+            - device 1 : TP = {2, 1} DP = {2, 1} EP = {1, 0}
+            - Comment: There are 2 engine instances and the tensors are sharded
+                across 2 decvices.
+
+            When TP = 2, DP = 2 and EP = False, the configuration on different
+                devices:
+
+            - device 0: TP = {4, 0} DP = {2, 0} EP = {1, 0}
+            - device 1: TP = {4, 1} DP = {2, 0} EP = {1, 0}
+            - device 2: TP = {4, 2} DP = {2, 1} EP = {1, 0}
+            - device 3: TP = {4, 3} DP = {2, 1} EP = {1, 0}
+            - Comment: There are 2 engine instances and the tensors are sharded
+                across 4 devices.
+
+            When, TP = 2, DP = 1 and EP = True, the configuration on different
+                devices:
+
+            - device 0: TP = {1, 0} DP = {1, 0} EP = {2, 0}
+            - device 1: TP = {1, 0} DP = {1, 0} EP = {2, 1}
+            - Comment: The experts are split between the 2 devices.
+
+            When, TP = 1, DP = 2 and EP = True, the configuration on different
+                devices:
+
+            - device 0: TP = {1, 0} DP = {2, 0} EP = {2, 0}
+            - device 1: TP = {1, 0} DP = {2, 1} EP = {2, 1}
+            - Comment: There are 2 engine instances and the experts are split
+                between the 2 devices.
+
+            When TP = 2, DP = 2 and EP = True, the configuration on different
+                devices:
+
+            - device 0: TP = {1, 0} DP = {2, 0} EP = {4, 0}
+            - device 1: TP = {1, 0} DP = {2, 0} EP = {4, 1}
+            - device 2: TP = {1, 0} DP = {2, 1} EP = {4, 2}
+            - device 3: TP = {1, 0} DP = {2, 1} EP = {4, 3}
+            - Comment: There are 2 engine instances and the experts are split
+                between the 4 devices.
+        """
+
+        use_ep = dp_size_ * tp_size_ > 1 and vllm_parallel_config.enable_expert_parallel
+
+        dp_size = dp_size_
+        dp_rank = get_dp_group().rank_in_group if dp_size > 1 else 0
+        tp_size, tp_rank = FusedMoEParallelConfig.flatten_tp_across_dp(
+            tp_size_, dp_size_, dp_rank
+        )
+
+        if not use_ep:
+            return FusedMoEParallelConfig(
+                tp_size=tp_size,
+                tp_rank=tp_rank,
+                dp_size=dp_size,
+                dp_rank=dp_rank,
+                ep_size=1,
+                ep_rank=0,
+                use_ep=False,
+                all2all_backend=vllm_parallel_config.all2all_backend,
+            )
+        # DP + EP / TP + EP / DP + TP + EP
+        assert use_ep
+        # In EP, each device owns a set of experts fully. There is no tensor
+        # parallel update tp_size, tp_rank, ep_size and ep_rank to reflect that.
+        ep_size = tp_size
+        ep_rank = tp_rank
+        return FusedMoEParallelConfig(
+            tp_size=1,
+            tp_rank=0,
+            dp_size=dp_size,
+            dp_rank=dp_rank,
+            ep_size=ep_size,
+            ep_rank=ep_rank,
+            use_ep=True,
+            all2all_backend=vllm_parallel_config.all2all_backend,
+        )
+
+
+# Adapted from pplx-kernels tests/all_to_all_utils.py
+@dataclass
+class FusedMoEConfig:
+    num_experts: int
+    experts_per_token: int
+    hidden_dim: int
+
+    num_local_experts: int
+    moe_parallel_config: FusedMoEParallelConfig
+
+    # The activation type.
+    in_dtype: torch.dtype
+
+    max_num_tokens: int = envs.VLLM_MOE_DP_CHUNK_SIZE
+
+    has_bias: bool = False
+
+    is_act_and_mul: bool = True
+
+    is_lora_enabled: bool = False
+
+    def __post_init__(self):
+        if self.dp_size > 1:
+            logger.debug_once(
+                "Using FusedMoEConfig::max_num_tokens=%d", self.max_num_tokens
+            )
+
+        assert self.max_num_tokens > 0
+
+    @property
+    def tp_size(self):
+        return self.moe_parallel_config.tp_size
+
+    @property
+    def dp_size(self):
+        return self.moe_parallel_config.dp_size
+
+    @property
+    def ep_size(self):
+        return self.moe_parallel_config.ep_size
+
+    @property
+    def tp_rank(self):
+        return self.moe_parallel_config.tp_rank
+
+    @property
+    def dp_rank(self):
+        return self.moe_parallel_config.dp_rank
+
+    @property
+    def ep_rank(self):
+        return self.moe_parallel_config.ep_rank
+
+    @property
+    def use_ep(self):
+        return self.moe_parallel_config.use_ep
+
+    @property
+    def use_pplx_kernels(self):
+        return self.moe_parallel_config.use_pplx_kernels
+
+    @property
+    def use_deepep_ht_kernels(self):
+        return self.moe_parallel_config.use_deepep_ht_kernels
+
+    @property
+    def use_deepep_ll_kernels(self):
+        return self.moe_parallel_config.use_deepep_ll_kernels
+
+    @property
+    def use_flashinfer_cutlass_kernels(self):
+        """
+        Whether to use FlashInfer cutlass kernels for NVFP4 MoE.
+        """
+        return (
+            envs.VLLM_USE_FLASHINFER_MOE_FP4
+            and has_flashinfer_cutlass_fused_moe()
+            and envs.VLLM_FLASHINFER_MOE_BACKEND == "throughput"
+        )

model_executor/layers/fused_moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json

@@ -0,0 +1,146 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "24": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "32": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 2
+  },
+  "48": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "64": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "96": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "128": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "256": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "512": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json

@@ -0,0 +1,146 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "24": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "32": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "48": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "64": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "96": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "128": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "256": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 4
+  },
+  "512": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json

@@ -0,0 +1,218 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "24": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "32": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "48": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "64": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "96": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "128": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "256": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "5120": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "9216": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "13312": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "17408": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "25600": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "33792": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "41984": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "50176": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "58368": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json

@@ -0,0 +1,218 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "24": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "32": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "48": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "64": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "96": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "128": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "256": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "5120": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "9216": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "13312": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "17408": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "25600": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "33792": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "41984": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "50176": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "58368": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=1792,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+  }

model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json

@@ -0,0 +1,218 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "24": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "32": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "48": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "64": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "96": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "128": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "256": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "5120": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "9216": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "13312": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "17408": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "25600": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "33792": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "41984": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "50176": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "58368": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json

@@ -0,0 +1,146 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "24": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "32": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "48": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "64": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "96": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "128": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "256": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json

@@ -0,0 +1,218 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "24": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "32": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "48": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "64": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "96": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "128": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "256": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "5120": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 4
+  },
+  "9216": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "13312": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "17408": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "25600": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 4
+  },
+  "33792": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "41984": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "50176": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "58368": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 3
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_H200,dtype=int8_w8a16.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json

@@ -0,0 +1,218 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "24": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "32": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "48": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "64": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "96": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "128": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "256": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "512": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 256,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "5120": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "9216": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "13312": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "17408": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "25600": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "33792": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "41984": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "50176": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "58368": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json

@@ -0,0 +1,218 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "24": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "32": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "48": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "64": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "96": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "128": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "256": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 4
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 8,
+    "num_stages": 4
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "5120": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "9216": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "13312": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "17408": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "25600": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "33792": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "41984": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "50176": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "58368": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json

@@ -0,0 +1,218 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "24": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "32": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "48": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "64": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "96": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "128": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "256": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "5120": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "9216": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "13312": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "17408": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "25600": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "33792": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "41984": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "50176": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "58368": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json

@@ -0,0 +1,218 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "16": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "24": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "32": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "48": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "64": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "96": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "128": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "256": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 8,
+    "num_stages": 4
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "5120": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "9216": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "13312": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "17408": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "25600": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "33792": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "41984": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "50176": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "58368": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 3
+  }
+}

model_executor/layers/fused_moe/configs/E=1,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=AMD_Instinct_MI300X.json

@@ -0,0 +1,200 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 1,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=NVIDIA_H100,dtype=fp8_w8a8.json

@@ -0,0 +1,123 @@
+{
+    "triton_version": "3.4.0",
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8192": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "16384": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8.json

@@ -0,0 +1,146 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "16": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "24": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "32": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "48": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "64": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "96": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "128": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "256": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  }
+}

model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=NVIDIA_H200.json

@@ -0,0 +1,146 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "4": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "8": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "16": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "24": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 4,
+    "num_stages": 2
+  },
+  "32": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "48": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "64": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "96": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "128": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 16,
+    "num_warps": 4,
+    "num_stages": 2
+  },
+  "256": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 64,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  }
+}

model_executor/layers/fused_moe/configs/E=128,N=1856,device_name=NVIDIA_H100_80GB_HBM3.json

@@ -0,0 +1,147 @@
+{
+    "triton_version": "3.5.0",
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=1856,device_name=NVIDIA_L40S.json

@@ -0,0 +1,147 @@
+{
+    "triton_version": "3.5.0",
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_A100-SXM4-80GB.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H100_80GB_HBM3.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H20-3e.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H20.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H200.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=352,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json

@@ -0,0 +1,122 @@
+{
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "8192": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16384": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20-3e.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H200.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=512,device_name=NVIDIA_H100_80GB_HBM3.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "48": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=704,device_name=NVIDIA_B200,dtype=fp8_w8a8.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 256,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=704,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json

@@ -0,0 +1,114 @@
+{
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "8192": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "16384": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=768,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=768,device_name=AMD_Instinct_MI308X.json

@@ -0,0 +1,213 @@
+{
+    "triton_version": "3.4.0",
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 1,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 4,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 1,
+        "num_stages": 2,
+        "waves_per_eu": 4,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 4,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 4,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 1,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 1,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 4,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 1,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "768": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 2,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 4,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H20.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H200.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=8960,device_name=NVIDIA_H100_80GB_HBM3,dtype=bf16.json

@@ -0,0 +1,82 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=8960,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json

@@ -0,0 +1,82 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "64": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=928,device_name=NVIDIA_H100_80GB_HBM3.json

@@ -0,0 +1,147 @@
+{
+    "triton_version": "3.5.0",
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 256,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 256,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=928,device_name=NVIDIA_L40S.json

@@ -0,0 +1,147 @@
+{
+    "triton_version": "3.5.0",
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

model_executor/layers/fused_moe/configs/E=128,N=96,device_name=NVIDIA_H20.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=AMD_Instinct_MI300X.json

@@ -0,0 +1,200 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 1
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 2,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0,
+        "matrix_instr_nonkdim": 16,
+        "kpack": 2
+    }
+}

model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_B200,dtype=fp8_w8a8.json

@@ -0,0 +1,147 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 256,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}
+

model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_B200.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}

model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_H100.json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 256,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}