Files
enginex-ascend-910-vllm/vllm_ascend/models/deepseek_v4.py
Sun Ruoxi 7f8a1b1f7a init v0.23.0
Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
2026-08-27 15:11:51 +08:00

1542 lines
63 KiB
Python

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# Copyright 2023 The vLLM team.
# Copyright 2023 DeepSeek-AI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import math
import typing
from collections.abc import Callable, Iterable
from itertools import islice
import torch
import torch.nn.functional as F
import torch_npu
from torch import nn
from transformers import DeepseekV2Config, DeepseekV3Config
from vllm._aiter_ops import rocm_aiter_ops
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ParallelConfig, VllmConfig
from vllm.distributed import (
get_ep_group,
get_pp_group,
get_tensor_model_parallel_rank,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_gather,
)
from vllm.model_executor.layers.activation import SiluAndMul, SiluAndMulWithClamp
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
MergedColumnParallelLinear,
ReplicatedLinear,
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead, VocabParallelEmbedding
from vllm.model_executor.model_loader.weight_utils import default_weight_loader, maybe_remap_kv_scale_name
from vllm.model_executor.models.interfaces import MixtureOfExperts, SupportsEagle, SupportsLoRA, SupportsPP
from vllm.model_executor.models.utils import (
PPMissingLayer,
is_pp_missing_parameter,
make_layers,
maybe_prefix,
sequence_parallel_chunk,
)
from vllm.models.deepseek_v4.attention import DeepseekV4IndexerCache # type: ignore[import-not-found,no-redef]
from vllm.models.deepseek_v4.compressor import CompressorStateCache # type: ignore[import-not-found,no-redef]
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs.deepseek_v4 import DeepseekV4Config
from vllm.v1.attention.backends.mla.sparse_swa import DeepseekV4SWACache as VllmDeepseekV4SWACache
from vllm.v1.kv_cache_interface import KVCacheSpec
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.core.kv_cache_interface import AscendSlidingWindowMLASpec
from vllm_ascend.ops.dsa import AscendDeepseekSparseAttention, DSAModules
from vllm_ascend.ops.rope_dsv4 import ComplexExpRotaryEmbedding
from vllm_ascend.ops.triton.mul_add import muls_add_triton
from vllm_ascend.utils import (
AscendDeviceType,
enable_dsa_cp,
extract_dsv4_layer_index,
get_ascend_device_type,
get_dsv4_compress_ratio,
vllm_version_is,
)
if not vllm_version_is("0.23.0"):
from vllm.model_executor.layers.fused_moe import fused_moe_make_expert_params_mapping
def _get_ascend_dsa_backend():
# Keep this lazy to avoid vLLM model-inspection circular imports.
from vllm_ascend.attention.dsa_v1 import AscendDSABackend
return AscendDSABackend
def _dsv4_block_sizes():
# Lazy import to avoid the circular import chain (layer -> dsa_v1 ->
# attention_v1 -> device_op) hit during vLLM subprocess model inspection.
from vllm_ascend.models.layer.attention.layer import DSV4_BLOCK_SIZES
return DSV4_BLOCK_SIZES
class AscendCompressorStateCache(CompressorStateCache):
def __init__(
self,
state_dim: int,
dtype: torch.dtype,
compress_ratio: int,
block_size: int,
prefix: str,
):
super().__init__(state_dim, dtype, compress_ratio, prefix)
self.compress_ratio = compress_ratio
self.block_size = block_size
def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
pads = _dsv4_block_sizes()[vllm_config.cache_config.block_size][1]
page_size_padded = pads[0] if self.state_dim == 2 * 256 and self.compress_ratio == 4 else pads[1]
return AscendSlidingWindowMLASpec(
block_size=self.block_size,
num_kv_heads=1,
head_size=self.state_dim,
dtype=self.dtype,
sliding_window=self.sliding_window,
alignment=None,
page_size_padded=page_size_padded,
)
def forward(self): ...
def get_attn_backend(self):
return _get_ascend_dsa_backend()
class AscendDeepseekV4IndexerCache(DeepseekV4IndexerCache):
def __init__(
self,
head_dim: int,
dtype: torch.dtype,
prefix: str,
cache_config: CacheConfig,
compress_ratio: int = 1,
):
super().__init__(head_dim, dtype, prefix, cache_config, compress_ratio)
def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
if get_ascend_device_type() in {AscendDeviceType.A5}:
self.dtype = torch.float8_e4m3fn
vllm_config.cache_config.cache_dtype = "float8_e4m3fn"
from vllm_ascend.core.kv_cache_interface import AscendMLAAttentionSpec
return AscendMLAAttentionSpec(
block_size=_dsv4_block_sizes()[vllm_config.cache_config.block_size][0][0],
num_kv_heads=1,
head_size=self.head_dim,
dtype=self.dtype,
model_version="deepseek_v4",
compress_ratio=self.compress_ratio,
cache_dtype_str=self.cache_config.cache_dtype,
scale_dim=1 if self.head_dim == 128 else 0,
scale_dtype=torch.float if get_ascend_device_type() in {AscendDeviceType.A5} else torch.float16,
)
def forward(self): ...
def get_attn_backend(self):
return _get_ascend_dsa_backend()
class AscendDeepseekV4SWACache(VllmDeepseekV4SWACache):
def __init__(
self,
head_dim: int,
window_size: int,
dtype: torch.dtype,
prefix: str,
cache_config: CacheConfig,
):
super().__init__(head_dim, window_size, torch.uint8, prefix, cache_config)
self.dtype = dtype
self.block_size = _dsv4_block_sizes()[cache_config.block_size][0][1]
def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec:
if get_ascend_device_type() in {AscendDeviceType.A5}:
self.dtype = torch.float8_e4m3fn
vllm_config.cache_config.cache_dtype = "float8_e4m3fn"
cached_head_size = self.head_dim + 128 if get_ascend_device_type() in {AscendDeviceType.A5} else self.head_dim
return AscendSlidingWindowMLASpec(
block_size=self.block_size,
num_kv_heads=1,
head_size=cached_head_size,
dtype=self.dtype,
sliding_window=self.window_size,
cache_dtype_str=self.cache_config.cache_dtype,
model_version="deepseek_v4",
alignment=None,
)
def forward(self): ...
def get_attn_backend(self):
return _get_ascend_dsa_backend()
def hadamard_transform_ref(x: torch.Tensor, scale=1.0):
from scipy.linalg import hadamard # type: ignore[import-untyped]
if hadamard is None:
raise ImportError("Please install scipy")
x_shape = x.shape
dim = x.shape[-1]
x = x.reshape(-1, dim)
log_dim = math.ceil(math.log2(dim))
dim_padded = 2**log_dim
if dim != dim_padded:
x = F.pad(x, (0, dim_padded - dim))
out = F.linear(x, torch.tensor(hadamard(dim_padded, dtype=float), dtype=x.dtype, device=x.device))
out = out * scale
return out[..., :dim].reshape(*x_shape)
def rotate_activation(x: torch.Tensor) -> torch.Tensor:
hidden_size = x.size(-1)
return hadamard_transform_ref(x, scale=hidden_size**-0.5)
def precompute_freqs_cis_cpu(dim, seqlen, original_seq_len, base, factor, beta_fast, beta_slow) -> torch.Tensor:
"""
Precomputes frequency-based complex exponential values for rotary positional embeddings.
Args:
args (ModelArgs): Model arguments containing positional embedding parameters.
Returns:
torch.Tensor: Precomputed complex exponential values for positional embeddings.
"""
def find_correction_dim(num_rotations, dim, base, max_seq_len):
return dim * math.log(max_seq_len / (num_rotations * 2 * math.pi)) / (2 * math.log(base))
def find_correction_range(low_rot, high_rot, dim, base, max_seq_len):
low = math.floor(find_correction_dim(low_rot, dim, base, max_seq_len))
high = math.ceil(find_correction_dim(high_rot, dim, base, max_seq_len))
return max(low, 0), min(high, dim - 1)
def linear_ramp_factor(min, max, dim):
if min == max:
max += 0.001
linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
ramp_func = torch.clamp(linear_func, 0, 1)
return ramp_func
freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
if original_seq_len > 0:
low, high = find_correction_range(beta_fast, beta_slow, dim, base, original_seq_len)
smooth = 1 - linear_ramp_factor(low, high, dim // 2)
freqs = freqs / factor * (1 - smooth) + freqs * smooth
t = torch.arange(seqlen)
freqs = torch.outer(t, freqs)
freqs_cis = torch.polar(torch.ones_like(freqs), freqs)
return freqs_cis
def apply_rotary_emb(x: torch.Tensor, freqs_cis: torch.Tensor, inverse: bool = False) -> torch.Tensor:
"""
Applies rotary positional embeddings to the input tensor.
Args:
x (torch.Tensor): Input tensor with positional embeddings to be applied.
freqs_cis (torch.Tensor): Precomputed complex exponential values for positional embeddings.
Returns:
torch.Tensor: Tensor with rotary embeddings applied.
"""
y = x
x = torch.view_as_complex(x.float().unflatten(-1, (-1, 2)))
if inverse:
freqs_cis = freqs_cis.conj()
if x.ndim == 3:
freqs_cis = freqs_cis.view(1, x.size(1), x.size(-1))
else:
freqs_cis = freqs_cis.view(1, x.size(1), 1, x.size(-1))
x = torch.view_as_real(x * freqs_cis.to(x.device)).flatten(-2)
y.copy_(x)
return y
def get_spec_layer_idx_from_weight_name(config: DeepseekV2Config | DeepseekV3Config, weight_name: str) -> int | None:
if weight_name.startswith("mtp."):
return 0
return None
class DeepseekV2MLP(nn.Module):
def __init__(
self,
hidden_size: int,
intermediate_size: int,
hidden_act: str,
swiglu_limit: float | None = None,
quant_config: QuantizationConfig | None = None,
reduce_results: bool = True,
is_sequence_parallel=False,
prefix: str = "",
) -> None:
super().__init__()
# If is_sequence_parallel, the input and output tensors are sharded
# across the ranks within the tp_group. In this case the weights are
# replicated and no collective ops are needed.
# Otherwise we use standard TP with an allreduce at the end.
self.gate_up_proj = MergedColumnParallelLinear(
hidden_size,
[intermediate_size] * 2,
bias=False,
quant_config=quant_config,
disable_tp=is_sequence_parallel,
prefix=f"{prefix}.gate_up_proj",
)
self.down_proj = RowParallelLinear(
intermediate_size,
hidden_size,
bias=False,
quant_config=quant_config,
reduce_results=reduce_results,
disable_tp=is_sequence_parallel,
prefix=f"{prefix}.down_proj",
)
if hidden_act != "silu":
raise ValueError(f"Unsupported activation: {hidden_act}. Only silu is supported for now.")
if swiglu_limit is not None:
self.act_fn = SiluAndMulWithClamp(swiglu_limit)
else:
self.act_fn = SiluAndMul()
def forward(self, x):
gate_up, _ = self.gate_up_proj(x)
x = self.act_fn(gate_up)
x, _ = self.down_proj(x)
return x
class DeepseekV4MoE(nn.Module):
def __init__(
self,
config: DeepseekV2Config | DeepseekV3Config | DeepseekV4Config,
parallel_config: ParallelConfig,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
is_draft_layer: bool = False,
):
super().__init__()
self.tp_size = get_tensor_model_parallel_world_size()
self.tp_rank = get_tensor_model_parallel_rank()
layer_idx = int(prefix.split(sep=".")[-2])
self.layer_idx = layer_idx
self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.5)
self.swiglu_limit = getattr(config, "swiglu_limit", None)
self.ep_group = get_ep_group().device_group
self.ep_rank = get_ep_group().rank_in_group
self.ep_size = self.ep_group.size()
self.n_routed_experts: int = config.n_routed_experts
self.n_shared_experts: int = config.n_shared_experts
self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe
if config.hidden_act != "silu":
raise ValueError(f"Unsupported activation: {config.hidden_act}. Only silu is supported for now.")
self.gate = ReplicatedLinear(
config.hidden_size, config.n_routed_experts, bias=False, quant_config=None, prefix=f"{prefix}.gate"
)
self.gate.precast_fp32_weight = True
# Load balancing settings.
eplb_config = parallel_config.eplb_config
self.enable_eplb = parallel_config.enable_eplb
self.n_redundant_experts = eplb_config.num_redundant_experts
self.n_logical_experts = self.n_routed_experts
self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
self.n_local_physical_experts = self.n_physical_experts // self.ep_size
self.physical_expert_start = self.ep_rank * self.n_local_physical_experts
self.physical_expert_end = self.physical_expert_start + self.n_local_physical_experts
self.is_rocm_aiter_moe_enabled = rocm_aiter_ops.is_fused_moe_enabled()
self.is_fusion_moe_shared_experts_enabled = rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
self.is_fusion_moe_shared_experts_enabled = getattr(get_ascend_config(), "mix_placement", False)
if config.n_shared_experts is None or self.is_fusion_moe_shared_experts_enabled:
self.shared_experts = None
else:
intermediate_size = config.moe_intermediate_size * config.n_shared_experts
self.shared_experts = DeepseekV2MLP(
hidden_size=config.hidden_size,
intermediate_size=intermediate_size,
hidden_act=config.hidden_act,
swiglu_limit=self.swiglu_limit,
quant_config=quant_config,
is_sequence_parallel=self.is_sequence_parallel,
reduce_results=False,
prefix=f"{prefix}.shared_experts",
)
self.hash = layer_idx < config.num_hash_layers and not is_draft_layer
if self.hash:
# Use zeros instead of empty to avoid garbage values causing
# invalid memory access in dummy mode (--load-format="dummy")
self.gate.tid2eid = nn.Parameter(
torch.zeros(
config.vocab_size,
config.num_experts_per_tok,
dtype=torch.int32,
),
requires_grad=False,
)
self.gate.e_score_correction_bias = None
else:
self.gate.tid2eid = None
self.gate.e_score_correction_bias = nn.Parameter(torch.empty(config.n_routed_experts, dtype=torch.float32))
self.experts = FusedMoE(
shared_experts=self.shared_experts,
gate=self.gate,
num_experts=config.n_routed_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
use_grouped_topk=True,
num_expert_group=getattr(config, "n_group", 1),
topk_group=getattr(config, "topk_group", 1),
prefix=f"{prefix}.experts",
scoring_func=getattr(config, "scoring_func", "softmax"),
# Keep scaling outside the router path so the order matches
# DeepSeek V4: normalize top-k weights, then scale routed output.
# AITER applies routed_scaling_factor internally.
routed_scaling_factor=self.routed_scaling_factor,
e_score_correction_bias=self.gate.e_score_correction_bias,
enable_eplb=self.enable_eplb,
num_redundant_experts=self.n_redundant_experts,
is_sequence_parallel=self.is_sequence_parallel,
n_shared_experts=config.n_shared_experts if self.is_fusion_moe_shared_experts_enabled else 0,
hash=layer_idx < config.num_hash_layers and not is_draft_layer,
tid2eid=self.gate.tid2eid,
)
def forward(self, hidden_states: torch.Tensor, input_ids=None) -> torch.Tensor:
num_tokens, hidden_dim = hidden_states.shape
hidden_states = hidden_states.view(-1, hidden_dim)
# Chunk the hidden states so they aren't replicated across TP ranks.
# This avoids duplicate computation in self.experts.
# TODO: We can replace the all_reduce at the end of attn with a
# reduce_scatter instead of chunking here.
if self.is_sequence_parallel:
hidden_states = sequence_parallel_chunk(hidden_states)
if self.experts.is_internal_router:
# In this case, the gate/router runs inside the FusedMoE class
fused_moe_out = self.experts(hidden_states=hidden_states, router_logits=hidden_states)
else:
# router_logits: (num_tokens, n_experts)
router_logits = F.linear(hidden_states.float(), self.gate.weight)
fused_moe_out = self.experts(hidden_states=hidden_states, router_logits=router_logits)
fused_moe_out_is_tuple = isinstance(fused_moe_out, tuple)
if fused_moe_out_is_tuple:
shared_output, final_hidden_states = fused_moe_out
if self.shared_experts is None:
assert shared_output is None
if hidden_states.dtype != torch.float16:
if not self.is_rocm_aiter_moe_enabled:
if self.shared_experts is not None:
assert shared_output is not None
final_hidden_states = muls_add_triton(
final_hidden_states, shared_output, self.routed_scaling_factor
)
else:
final_hidden_states *= self.routed_scaling_factor
elif self.shared_experts is not None:
assert shared_output is not None
final_hidden_states = muls_add_triton(
shared_output, final_hidden_states, 1.0 / self.routed_scaling_factor
)
else:
final_hidden_states = fused_moe_out
if self.is_sequence_parallel:
final_hidden_states = tensor_model_parallel_all_gather(final_hidden_states, 0)
final_hidden_states = final_hidden_states[:num_tokens]
elif self.tp_size > 1 and fused_moe_out_is_tuple:
# Legacy tuple outputs are reduced here. Tensor outputs from the
# upstream MoERunner have already gone through its final reduction.
final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(final_hidden_states)
return final_hidden_states.view(num_tokens, hidden_dim)
def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
import math
if scale <= 1:
return 1.0
return 0.1 * mscale * math.log(scale) + 1.0
def _get_llama_4_scaling(
original_max_position_embeddings: int, scaling_beta: float, positions: torch.Tensor
) -> torch.Tensor:
scaling = 1 + scaling_beta * torch.log(1 + torch.floor(positions / original_max_position_embeddings))
# Broadcast over num_heads and head_dim
return scaling[..., None, None]
class Indexer(nn.Module):
def __init__(
self,
vllm_config: VllmConfig,
config: DeepseekV2Config | DeepseekV3Config | DeepseekV4Config,
compress_ratio: int,
quant_config: QuantizationConfig | None,
cache_config: CacheConfig | None,
prefix: str = "",
):
super().__init__()
self.vllm_config = vllm_config
self.config = config
self.n_heads = config.index_n_heads
self.head_dim = config.index_head_dim
self.rope_head_dim = config.qk_rope_head_dim
self.index_topk = config.index_topk
self.q_lora_rank = config.q_lora_rank
self.softmax_scale = self.head_dim**-0.5
self.compress_ratio = compress_ratio
self.wq_b = ReplicatedLinear(
self.q_lora_rank,
self.n_heads * self.head_dim,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.wq_b",
return_bias=False,
)
self.weights_proj = ReplicatedLinear(
config.hidden_size,
self.n_heads,
bias=False,
quant_config=None,
prefix=f"{prefix}.weights_proj",
return_bias=False,
)
ascend_device_type = get_ascend_device_type()
k_dtype = torch.float8_e4m3fn if ascend_device_type == AscendDeviceType.A5 else torch.int8
if self.compress_ratio == 4:
# TODO(cmq): change the dtype of cache
self.k_cache = AscendDeepseekV4IndexerCache(
head_dim=self.head_dim,
dtype=k_dtype,
prefix=f"{prefix}.k_cache",
cache_config=cache_config,
compress_ratio=self.compress_ratio,
)
self.compressor = None
if self.compress_ratio > 1:
self.compressor = Compressor(
vllm_config,
config,
self.compress_ratio,
head_dim=self.head_dim,
rotate=True,
quant_config=quant_config,
cache_config=cache_config,
prefix=f"{prefix}.compressor",
) # Compressor(4, 128)
def forward(self, hidden_states: torch.Tensor, qr: torch.Tensor, positions, rotary_emb) -> torch.Tensor:
return
class Compressor(nn.Module):
def __init__(
self,
vllm_config: VllmConfig,
config: DeepseekV2Config | DeepseekV3Config | DeepseekV4Config,
compress_ratio: int = 4,
head_dim: int = 512,
rotate: bool = False,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
):
super().__init__()
self.vllm_config = vllm_config
self.config = config
self.dim = config.hidden_size
self.head_dim = head_dim
self.rope_head_dim = config.qk_rope_head_dim
self.nope_head_dim = head_dim - config.qk_rope_head_dim
self.compress_ratio = compress_ratio
self.overlap = compress_ratio == 4
self.rotate = rotate
self.norm_eps = config.rms_norm_eps
self.coff = 1 + self.overlap
self.ape = nn.Parameter(torch.empty(compress_ratio, self.coff * self.head_dim, dtype=torch.float32))
self.wkv = ReplicatedLinear(
self.dim,
self.coff * self.head_dim,
bias=False,
quant_config=None if get_ascend_device_type() in {AscendDeviceType.A5} else quant_config,
prefix=f"{prefix}.wkv",
return_bias=False,
)
self.wgate = ReplicatedLinear(
self.dim,
self.coff * self.head_dim,
bias=False,
quant_config=None if get_ascend_device_type() in {AscendDeviceType.A5} else quant_config,
prefix=f"{prefix}.wgate",
return_bias=False,
)
# A5 compressor kernel needs float for norm_weight input
norm_dtype = torch.float32 if get_ascend_device_type() == AscendDeviceType.A5 else None
self.norm = RMSNorm(self.head_dim, config.rms_norm_eps, dtype=norm_dtype)
state_dtype = torch.float32
# TODO(zyj): change following codes if block_size is configurable & refactor the magic numbers
if compress_ratio == 4:
self.state_cache = AscendCompressorStateCache(
state_dim=2 * self.coff * self.head_dim, # kv_state + score_state
dtype=state_dtype,
compress_ratio=compress_ratio,
prefix=f"{prefix}.state_cache",
block_size=_dsv4_block_sizes()[cache_config.block_size][0][2], # type: ignore[union-attr]
)
elif compress_ratio == 128:
self.state_cache = AscendCompressorStateCache(
state_dim=2 * self.head_dim, # kv_state + score_state
dtype=state_dtype,
compress_ratio=compress_ratio,
prefix=f"{prefix}.state_cache",
block_size=_dsv4_block_sizes()[cache_config.block_size][0][3], # type: ignore[union-attr]
)
else:
raise ValueError(
f"Only support compress_ratio in [4, 128]. Got unsupported compress_ratio: {compress_ratio}"
)
def overlap_transform(self, tensor: torch.Tensor, value=0):
b, s, _, _ = tensor.size()
ratio, d = self.compress_ratio, self.head_dim
new_tensor = tensor.new_full((b, s, 2 * ratio, d), value)
new_tensor[:, :, ratio:] = tensor[:, :, :, d:]
new_tensor[:, 1:, :ratio] = tensor[:, :-1, :, :d]
return new_tensor
def forward(
self,
x: torch.Tensor,
start_pos: int,
cos: torch.Tensor,
sin: torch.Tensor,
) -> torch.Tensor:
pass
def rope_single(
self,
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
inverse: bool = False,
) -> torch.Tensor:
dtype = x.dtype
if inverse:
sin = sin * -1
tnd_layout = 1
if len(x.shape) == 3:
num_tokens, num_heads, rotary_dim = x.shape
else:
tnd_layout = 0
_, num_tokens, num_heads, rotary_dim = x.shape
x_rot = torch_npu.npu_rotary_mul(
x.reshape(num_tokens, num_heads, 1, rotary_dim).to(torch.float32), cos, sin, rotary_mode="interleave"
)
if tnd_layout:
x = x_rot.reshape(num_tokens, -1, rotary_dim)
else:
x = x_rot.reshape(1, num_tokens, -1, rotary_dim)
return x.to(dtype)
class DeepseekV4Attention(nn.Module):
def __init__(
self,
vllm_config: VllmConfig,
config: DeepseekV2Config | DeepseekV3Config | DeepseekV4Config,
max_position_embeddings: int = 0,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
topk_indices_buffer: torch.Tensor | None = None,
) -> None:
super().__init__()
layer_idx = int(prefix.split(sep=".")[-2])
self.layer_idx = layer_idx
config_layer_idx = extract_dsv4_layer_index(config, prefix)
tp_size = get_tensor_model_parallel_world_size()
self.dim = config.hidden_size
self.n_heads = config.num_attention_heads
self.n_local_heads = config.num_attention_heads // tp_size
self.q_lora_rank = config.q_lora_rank
self.o_lora_rank = config.o_lora_rank
self.head_dim = config.head_dim
self.rope_head_dim = config.qk_rope_head_dim
self.nope_head_dim = config.head_dim - config.qk_rope_head_dim
self.n_groups = config.o_groups
self.n_local_groups = self.n_groups // tp_size
self.window_size = config.sliding_window
self.eps = config.rms_norm_eps
self.norm_eps = config.rms_norm_eps
self.scale = self.head_dim**-0.5
self.enable_dsa_cp = enable_dsa_cp()
attn_sink_heads = self.n_heads if self.enable_dsa_cp else self.n_local_heads
self.attn_sink = nn.Parameter(torch.empty(attn_sink_heads, dtype=torch.float32))
self.wq_a = ReplicatedLinear(
self.dim,
self.q_lora_rank,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.wq_a",
return_bias=False,
)
self.q_norm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
self.q_norm_without_weight = RMSNorm(self.head_dim, eps=config.rms_norm_eps, has_weight=False)
wq_b_cls = ReplicatedLinear if self.enable_dsa_cp else ColumnParallelLinear
self.wq_b = wq_b_cls(
self.q_lora_rank,
self.n_heads * self.head_dim,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.wq_b",
return_bias=False,
)
self.wkv = ReplicatedLinear(
self.dim,
self.head_dim,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.wkv",
return_bias=False,
)
self.kv_norm = RMSNorm(self.head_dim, self.norm_eps)
self.wo_a = ColumnParallelLinear(
self.n_heads * self.head_dim // self.n_groups,
self.n_groups * config.o_lora_rank,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.wo_a",
return_bias=False,
)
self.wo_b = RowParallelLinear(
self.n_groups * config.o_lora_rank,
self.dim,
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.wo_b",
return_bias=False,
)
self.compress_ratio = get_dsv4_compress_ratio(config, config_layer_idx)
if self.compress_ratio > 1:
config.rope_parameters["rope_theta"] = config.compress_rope_theta
rope_groups = ["default", f"c{self.compress_ratio}"]
else:
config.rope_parameters["rope_theta"] = config.rope_theta
rope_groups = ["default"]
self.rotary_emb = ComplexExpRotaryEmbedding(
vllm_config=vllm_config,
layername=f"{prefix}.attn",
head_size=self.rope_head_dim,
rotary_dim=self.rope_head_dim,
max_position_embeddings=max_position_embeddings,
is_neox_style=False,
scaling_factor=config.rope_parameters["factor"],
base=config.rope_parameters["rope_theta"],
beta_fast=config.rope_parameters["beta_fast"],
beta_slow=config.rope_parameters["beta_slow"],
rope_groups=rope_groups,
)
self.compressor: Compressor | None = None
self.indexer: Indexer | None = None
if self.compress_ratio > 1:
self.compressor = Compressor(
vllm_config,
config,
self.compress_ratio,
head_dim=self.head_dim,
quant_config=quant_config,
cache_config=cache_config,
prefix=f"{prefix}.compressor",
) # Compressor(4, 128)
if self.compress_ratio == 4:
self.indexer = Indexer(
vllm_config,
config,
self.compress_ratio,
quant_config=quant_config,
cache_config=cache_config,
prefix=f"{prefix}.indexer",
)
# IndexCache: decide whether this layer reuses topk from a previous
# indexer-bearing layer. Refer: https://arxiv.org/abs/2603.12201
# Only meaningful when this layer actually owns an Indexer (c4) and
# IndexCache is enabled via hf-overrides. MTP layers are excluded
# because spec_decode shares topk_indices_buffer at the model level
# only, leaving impl-level references stale.
skip_topk = False
if self.compress_ratio == 4 and getattr(config, "use_index_cache", False) and ".mtp." not in prefix:
compress_ratios = getattr(config, "compress_ratios", None) or []
indexer_seq_idx = sum(1 for r in compress_ratios[:config_layer_idx] if r == 4)
pattern = getattr(config, "index_topk_pattern", None)
freq = getattr(config, "index_topk_freq", 1)
if pattern is None:
skip_topk = max(indexer_seq_idx - 1, 0) % freq != 0
else:
assert pattern[0] == "F", "index_topk_pattern must start with 'F'"
if 0 <= indexer_seq_idx < len(pattern):
skip_topk = pattern[indexer_seq_idx] == "S"
ascend_device_type = get_ascend_device_type()
k_dtype = torch.float8_e4m3fn if ascend_device_type == AscendDeviceType.A5 else torch.bfloat16
swa_cache_layer = AscendDeepseekV4SWACache(
head_dim=self.head_dim,
window_size=self.window_size,
dtype=k_dtype,
prefix=f"{prefix}.swa_cache",
cache_config=cache_config,
)
dsa_modules = DSAModules(
wq_a=self.wq_a,
q_norm=self.q_norm,
q_norm_without_weight=self.q_norm_without_weight,
wq_b=self.wq_b,
wkv=self.wkv,
kv_norm=self.kv_norm,
wo_a=self.wo_a,
wo_b=self.wo_b,
attn_sink=self.attn_sink,
indexer=self.indexer,
compressor=self.compressor,
swa_cache_layer=swa_cache_layer,
topk_indices_buffer=topk_indices_buffer,
skip_topk=skip_topk,
)
self.dsa_attn = AscendDeepseekSparseAttention(
dim=self.dim,
n_heads=self.n_heads,
scale=self.scale,
n_local_heads=self.n_local_heads,
q_lora_rank=self.q_lora_rank,
o_lora_rank=self.o_lora_rank,
head_dim=self.head_dim,
rope_head_dim=self.rope_head_dim,
nope_head_dim=self.nope_head_dim,
eps=self.eps,
n_groups=self.n_groups,
n_local_groups=self.n_local_groups,
window_size=self.window_size,
compress_ratio=self.compress_ratio,
dsa_modules=dsa_modules,
cache_config=cache_config,
quant_config=quant_config,
# prefix=f'{prefix}.attn',
prefix=f"{prefix}",
)
def forward(
self,
positions: torch.Tensor,
hidden_states: torch.Tensor,
llama_4_scaling: torch.Tensor | None,
) -> torch.Tensor:
return self.dsa_attn(positions, hidden_states, llama_4_scaling)
class DeepseekV2DecoderLayer(nn.Module):
def __init__(
self,
vllm_config: VllmConfig,
prefix: str,
config: DeepseekV2Config | None = None,
topk_indices_buffer: torch.Tensor | None = None,
is_draft_layer: bool = False,
) -> None:
super().__init__()
if config is None:
config = vllm_config.model_config.hf_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
parallel_config = vllm_config.parallel_config
self.hidden_size = config.hidden_size
max_position_embeddings = config.rope_parameters["original_max_position_embeddings"]
# DecoderLayers are created with `make_layers` which passes the prefix
# with the layer's index.
layer_idx = int(prefix.split(sep=".")[-1])
self.layer_idx = layer_idx
self.norm_eps = config.rms_norm_eps
attn_cls = DeepseekV4Attention
self.self_attn = attn_cls(
vllm_config=vllm_config,
config=config,
max_position_embeddings=max_position_embeddings,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.self_attn",
topk_indices_buffer=topk_indices_buffer,
)
self.mlp = DeepseekV4MoE(
config=config,
parallel_config=parallel_config,
quant_config=quant_config,
prefix=f"{prefix}.mlp",
is_draft_layer=is_draft_layer,
)
self.input_layernorm = RMSNorm(config.hidden_size, eps=self.norm_eps)
self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=self.norm_eps)
self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0)
self.hc_mult = hc_mult = config.hc_mult
self.hc_sinkhorn_iters = config.hc_sinkhorn_iters
self.hc_eps = config.hc_eps
mix_hc = (2 + hc_mult) * hc_mult
hc_dim = hc_mult * config.hidden_size
self.hc_attn_fn = nn.Parameter(torch.empty(mix_hc, hc_dim, dtype=torch.float32))
self.hc_ffn_fn = nn.Parameter(torch.empty(mix_hc, hc_dim, dtype=torch.float32))
self.hc_attn_base = nn.Parameter(torch.empty(mix_hc, dtype=torch.float32))
self.hc_ffn_base = nn.Parameter(torch.empty(mix_hc, dtype=torch.float32))
self.hc_attn_scale = nn.Parameter(torch.empty(3, dtype=torch.float32))
self.hc_ffn_scale = nn.Parameter(torch.empty(3, dtype=torch.float32))
def hc_pre(self, x: torch.Tensor, hc_fn: torch.Tensor, hc_scale: torch.Tensor, hc_base: torch.Tensor):
y = torch.ops._C_ascend.npu_hc_pre_v2(
x, hc_fn, hc_scale, hc_base, self.hc_mult, self.hc_sinkhorn_iters, self.norm_eps, self.hc_eps
)
return y
def hc_post(self, x: torch.Tensor, residual: torch.Tensor, post: torch.Tensor, comb: torch.Tensor):
y = torch.ops._C_ascend.npu_hc_post(
x.unsqueeze(dim=0), residual.unsqueeze(dim=0), post.unsqueeze(dim=0), comb.unsqueeze(dim=0)
)
return y.squeeze(dim=0)
def forward(
self,
positions: torch.Tensor,
hidden_states: torch.Tensor,
residual: torch.Tensor | None,
llama_4_scaling: torch.Tensor | None = None,
) -> torch.Tensor:
residual = hidden_states.clone()
hidden_states, post, comb = self.hc_pre(hidden_states, self.hc_attn_fn, self.hc_attn_scale, self.hc_attn_base)
hidden_states = self.input_layernorm(hidden_states)
attn_kwargs = {"positions": positions, "hidden_states": hidden_states, "llama_4_scaling": llama_4_scaling}
hidden_states = self.self_attn(**attn_kwargs)
hidden_states = self.hc_post(hidden_states, residual, post, comb)
residual = hidden_states.clone()
hidden_states, post, comb = self.hc_pre(hidden_states, self.hc_ffn_fn, self.hc_ffn_scale, self.hc_ffn_base)
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = self.hc_post(hidden_states, residual, post, comb)
return hidden_states, residual
@support_torch_compile
class DeepseekV4Model(nn.Module):
fall_back_to_pt_during_load = False
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
self.config = config
self.device = current_platform.device_type
self.vocab_size = config.vocab_size
self.is_v32 = hasattr(config, "index_topk")
if self.is_v32:
topk_tokens = config.index_topk
topk_indices_buffer = torch.empty(
vllm_config.scheduler_config.max_num_batched_tokens,
topk_tokens,
dtype=torch.int32,
device=self.device,
)
else:
topk_indices_buffer = None
# Expose at model level so spec_decode/llm_base_proposer can share
# this buffer with the MTP draft via attribute replacement.
self.topk_indices_buffer = topk_indices_buffer
if get_pp_group().is_first_rank:
self.embed_tokens = VocabParallelEmbedding(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=f"{prefix}.embed_tokens",
)
else:
self.embed_tokens = PPMissingLayer()
self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers,
lambda prefix: DeepseekV2DecoderLayer(vllm_config, prefix, topk_indices_buffer=topk_indices_buffer),
prefix=f"{prefix}.layers",
)
if get_pp_group().is_last_rank:
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
else:
self.norm = PPMissingLayer()
def make_empty_intermediate_tensors(
batch_size: int,
dtype: torch.dtype,
device: torch.device,
) -> IntermediateTensors:
return IntermediateTensors(
{
"hidden_states": torch.zeros(
(batch_size, self.hc_mult, config.hidden_size),
dtype=dtype,
device=device,
),
}
)
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors
self.norm_eps = config.rms_norm_eps
self.hc_eps = config.hc_eps
self.hc_mult = hc_mult = config.hc_mult
hc_dim = hc_mult * config.hidden_size
self.hc_head_fn = nn.Parameter(torch.empty(hc_mult, hc_dim, dtype=torch.float32))
self.hc_head_base = nn.Parameter(torch.empty(hc_mult, dtype=torch.float32))
self.hc_head_scale = nn.Parameter(torch.empty(1, dtype=torch.float32))
# Pre-hc_head residual stream buffer for the MTP draft. Stable
# address (outside the cudagraph pool) so the copy_ in forward()
# refreshes it correctly across captured shapes.
self._mtp_hidden_buffer = torch.empty(
vllm_config.scheduler_config.max_num_batched_tokens,
hc_dim,
dtype=vllm_config.model_config.dtype,
device=self.device,
)
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.embed_tokens(input_ids)
def hc_head(self, x: torch.Tensor, hc_fn: torch.Tensor, hc_scale: torch.Tensor, hc_base: torch.Tensor):
shape, dtype = x.size(), x.dtype
x = x.flatten(1).float()
rsqrt = torch.rsqrt(x.square().mean(-1, keepdim=True) + self.norm_eps)
mixes = torch.nn.functional.linear(x, hc_fn) * rsqrt
pre = torch.sigmoid(mixes * hc_scale + hc_base) + self.hc_eps
y = torch.sum(pre.unsqueeze(-1) * x.view(shape), dim=1)
return y.to(dtype)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors: IntermediateTensors | None,
inputs_embeds: torch.Tensor | None = None,
) -> torch.Tensor | IntermediateTensors:
if get_pp_group().is_first_rank:
if inputs_embeds is not None:
hidden_states = inputs_embeds
else:
hidden_states = self.embed_input_ids(input_ids)
residual = None
else:
assert intermediate_tensors is not None
hidden_states = intermediate_tensors["hidden_states"]
residual = None
# Compute llama 4 scaling once per forward pass if enabled
llama_4_scaling_config = None
llama_4_scaling: torch.Tensor | None
if llama_4_scaling_config is not None:
llama_4_scaling = _get_llama_4_scaling(
original_max_position_embeddings=llama_4_scaling_config["original_max_position_embeddings"],
scaling_beta=llama_4_scaling_config["beta"],
positions=positions,
)
else:
llama_4_scaling = None
if get_pp_group().is_first_rank:
hidden_states = hidden_states.unsqueeze(1).repeat(1, self.hc_mult, 1) # (b, s, h) -> (b, s, c, h)
for layer in islice(self.layers, self.start_layer, self.end_layer):
hidden_states, residual = layer(positions, hidden_states, residual, llama_4_scaling)
# Stash pre-hc_head residual for the MTP draft (captured copy_).
# When FlashComm1 (sequence parallelism) is enabled, tokens are
# partitioned across TP ranks via reduce_scatter in each layer's
# row-parallel output projection. We must all_gather here so the
# MTP layers receive the full token set — otherwise only rank 0's
# partition is valid and the rest of the buffer holds stale data,
# leading to NaN values and low acceptance rate.
from vllm_ascend.ascend_forward_context import get_forward_context
forward_ctx = get_forward_context()
if forward_ctx is not None and forward_ctx.flash_comm_v1_enabled:
h_states_flat = tensor_model_parallel_all_gather(hidden_states.flatten(1), dim=0)
pad_size = forward_ctx.pad_size
if pad_size > 0:
h_states_flat = h_states_flat[:-pad_size]
num_tokens = h_states_flat.shape[0]
self._mtp_hidden_buffer[:num_tokens].copy_(h_states_flat)
else:
num_tokens = hidden_states.shape[0]
self._mtp_hidden_buffer[:num_tokens].copy_(hidden_states.flatten(1))
if not get_pp_group().is_last_rank:
return IntermediateTensors(
{
"hidden_states": hidden_states,
}
)
hidden_states = self.hc_head(hidden_states, self.hc_head_fn, self.hc_head_scale, self.hc_head_base)
hidden_states = self.norm(hidden_states)
return hidden_states
class DeepseekV2MixtureOfExperts(MixtureOfExperts):
moe_mlp_layers: list[DeepseekV4MoE]
"""
List of MoE MLP layers in the model.
"""
def extract_moe_parameters(self, example_moe: DeepseekV4MoE | None):
if example_moe is None:
self.num_moe_layers = 0
self.num_expert_groups = 0
self.num_logical_experts = 0
self.num_physical_experts = 0
self.num_local_physical_experts = 0
self.num_routed_experts = 0
self.num_shared_experts = 0
self.num_redundant_experts = 0
else:
self.num_logical_experts = example_moe.n_logical_experts
self.num_physical_experts = example_moe.n_physical_experts
self.num_local_physical_experts = example_moe.n_local_physical_experts
self.num_routed_experts = example_moe.n_routed_experts
self.num_shared_experts = example_moe.n_shared_experts
self.num_redundant_experts = example_moe.n_redundant_experts
def update_physical_experts_metadata(
self,
num_physical_experts: int,
num_local_physical_experts: int,
) -> None:
assert self.num_local_physical_experts == num_local_physical_experts
self.num_physical_experts = num_physical_experts
self.num_local_physical_experts = num_local_physical_experts
self.num_redundant_experts = num_physical_experts - self.num_logical_experts
for moe in self.moe_mlp_layers:
moe.n_local_physical_experts = num_local_physical_experts
moe.n_physical_experts = num_physical_experts
moe.n_redundant_experts = self.num_redundant_experts
moe.experts.update_expert_map()
class AscendDeepseekV4ForCausalLM(nn.Module, SupportsPP, DeepseekV2MixtureOfExperts, SupportsLoRA, SupportsEagle):
packed_modules_mapping = {
"gate_up_proj": ["gate_proj", "up_proj"],
}
model_cls = DeepseekV4Model
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
self.config = config
self.quant_config = quant_config
self.model = self.model_cls(vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model"))
if get_pp_group().is_last_rank:
self.lm_head = ParallelLMHead(
config.vocab_size,
config.hidden_size,
quant_config=quant_config,
prefix=maybe_prefix(prefix, "lm_head"),
)
else:
self.lm_head = PPMissingLayer()
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = self.model.make_empty_intermediate_tensors
# Set MoE hyperparameters
self.num_moe_layers = self.config.num_hidden_layers
self.set_moe_parameters()
def set_moe_parameters(self):
self.expert_weights = []
self.num_expert_groups = getattr(self.config, "n_group", 1)
self.moe_layers = []
self.moe_mlp_layers = []
example_moe = None
for layer in self.model.layers:
if isinstance(layer, PPMissingLayer):
continue
assert isinstance(layer, DeepseekV2DecoderLayer)
if isinstance(layer.mlp, DeepseekV4MoE):
# Pick last one layer since the first ones may be dense layers.
example_moe = layer.mlp
self.moe_mlp_layers.append(layer.mlp)
self.moe_layers.append(layer.mlp.experts)
self.extract_moe_parameters(example_moe)
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.model.embed_input_ids(input_ids)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors: IntermediateTensors | None = None,
inputs_embeds: torch.Tensor | None = None,
) -> torch.Tensor | IntermediateTensors:
hidden_states = self.model(input_ids, positions, intermediate_tensors, inputs_embeds)
return hidden_states
def compute_logits(
self,
hidden_states: torch.Tensor,
) -> torch.Tensor | None:
logits = self.logits_processor(self.lm_head, hidden_states)
return logits
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
# Params for weights, fp8 weight scales, fp8 activation scales
# (param_name, weight_name, expert_id, shard_id)
if vllm_version_is("0.23.0"):
return FusedMoE.make_expert_params_mapping(
self.model,
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.n_routed_experts
+ (self.config.n_shared_experts if getattr(get_ascend_config(), "mix_placement", False) else 0),
num_redundant_experts=0,
)
else:
return fused_moe_make_expert_params_mapping(
self.model,
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.n_routed_experts
+ (self.config.n_shared_experts if getattr(get_ascend_config(), "mix_placement", False) else 0),
num_redundant_experts=0,
)
def get_mtp_target_hidden_states(self) -> torch.Tensor | None:
"""Pre-hc_head residual stream buffer (max_num_batched_tokens,
hc_mult * hidden_size) for the MTP draft model. Populated by
forward(); valid after each target step."""
return getattr(self.model, "_mtp_hidden_buffer", None)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
rocm_aiter_moe_shared_expert_enabled = rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
rocm_aiter_moe_shared_expert_enabled = getattr(get_ascend_config(), "mix_placement", False)
stacked_params_mapping = [
("gate_up_proj", "gate_proj", 0),
("gate_up_proj", "up_proj", 1),
]
# Params for weights, fp8 weight scales, fp8 activation scales
# (param_name, weight_name, expert_id, shard_id)
if vllm_version_is("0.23.0"):
expert_params_mapping = FusedMoE.make_expert_params_mapping(
self.model,
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.n_routed_experts
+ (self.config.n_shared_experts if rocm_aiter_moe_shared_expert_enabled else 0),
num_redundant_experts=self.num_redundant_experts,
)
else:
expert_params_mapping = fused_moe_make_expert_params_mapping(
self.model,
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",
ckpt_up_proj_name="up_proj",
num_experts=self.config.n_routed_experts
+ (self.config.n_shared_experts if rocm_aiter_moe_shared_expert_enabled else 0),
num_redundant_experts=self.num_redundant_experts,
)
params_dict = dict(self.named_parameters())
loaded_params: set[str] = set()
tp_rank = get_tensor_model_parallel_rank()
tp_size = get_tensor_model_parallel_world_size()
# Attention heads per rank
heads_per_rank = self.config.num_attention_heads // tp_size
head_start = tp_rank * heads_per_rank
for name, loaded_weight in weights:
spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
if spec_layer is not None:
continue # skip spec decode layers for main model
# TODO:
if not name.startswith("model"):
name = f"model.{name}"
if ".w1." in name:
name = name.replace(".w1.", ".gate_proj.")
if ".w2." in name:
name = name.replace(".w2.", ".down_proj.")
if ".w3." in name:
name = name.replace(".w3.", ".up_proj.")
if "model.head." in name and "model.lm_head." not in name:
name = name.replace("model.head.", "lm_head.")
if "model.lm_head." in name:
name = name.replace("model.lm_head.", "lm_head.")
if "embed." in name and "embed_token." not in name:
name = name.replace("embed.", "embed_tokens.")
if "attn" in name and "self_attn" not in name:
name = name.replace(".attn.", ".self_attn.")
if ".ffn." in name:
name = name.replace(".ffn.", ".mlp.")
if ".ffn_norm." in name:
name = name.replace(".ffn_norm.", ".post_attention_layernorm.")
if ".attn_norm." in name:
name = name.replace(".attn_norm.", ".input_layernorm.")
if name.endswith(".scale"):
name = name.replace(".scale", ".weight_scale")
if "rotary_emb.inv_freq" in name:
continue
if ".gate.bias" in name:
name = name.replace(".gate.bias", ".gate.e_score_correction_bias")
if "sink" in name:
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
if enable_dsa_cp():
param.data.copy_(loaded_weight)
else:
# Handle attention sinks (distributed across ranks)
narrow_weight = loaded_weight.narrow(0, head_start, heads_per_rank)
param.data.copy_(narrow_weight)
loaded_params.add(name)
continue
is_fusion_moe_shared_experts_layer = rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
for param_name, weight_name, shard_id in stacked_params_mapping:
# Skip non-stacked layers and experts (experts handled below).
if weight_name not in name:
continue
# We have mlp.experts[0].gate_proj in the checkpoint.
# Since we handle the experts below in expert_params_mapping,
# we need to skip here BEFORE we update the name, otherwise
# name will be updated to mlp.experts[0].gate_up_proj, which
# will then be updated below in expert_params_mapping
# for mlp.experts[0].gate_gate_up_proj, which breaks load.
if ("mlp.experts." in name) and name not in params_dict:
continue
if is_fusion_moe_shared_experts_layer:
continue
name_mapped = name.replace(weight_name, param_name)
# QKV fusion is optional, fall back to normal
# weight loading if it's not enabled
# if go with fusion option, then update name
if (param_name == "fused_qkv_a_proj") and name_mapped not in params_dict:
continue
else:
name = name_mapped
# Skip loading extra bias for GPTQ models.
if name.endswith(".bias") and name not in params_dict:
continue
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
weight_loader = param.weight_loader
weight_loader(param, loaded_weight, shard_id)
break
else:
is_expert_weight = False
# Special handling: when AITER fusion_shared_experts is enabled,
# checkpoints may provide a single widened shared_experts tensor
# without explicit expert indices
# (e.g. ...mlp.shared_experts.gate_proj.weight).
# For models with multiple shared experts, split that tensor
# evenly into per-shared-expert slices and load them into
# appended expert slots mlp.experts.{n_routed_experts + j}.*
# accordingly.
num_chunks = 1
if is_fusion_moe_shared_experts_layer:
num_chunks = getattr(self.config, "n_shared_experts", 1) or 1
# Determine split axis based on op type
# gate/up: ColumnParallel → split along dim 0
# down: RowParallel → split along dim 1
split_dim = 1 if "down_proj.weight" in name else 0
total = loaded_weight.shape[split_dim]
assert total % num_chunks == 0, (
f"Shared expert weight dim {total} not divisible by num_chunks {num_chunks}"
)
chunk_size = total // num_chunks
for j in range(num_chunks):
chunk_name = name
weight_to_load = loaded_weight
if is_fusion_moe_shared_experts_layer:
if split_dim == 0:
weight_to_load = loaded_weight[j * chunk_size : (j + 1) * chunk_size, :]
else:
weight_to_load = loaded_weight[:, j * chunk_size : (j + 1) * chunk_size]
# Synthesize an expert-style name so expert mapping
# can route it
chunk_name = name.replace(
"mlp.shared_experts",
f"mlp.experts.{self.config.n_routed_experts + j}",
)
# Use expert_params_mapping to locate the destination
# param and delegate to its expert-aware weight_loader
# with expert_id.
for mapping in expert_params_mapping:
param_name, weight_name, expert_id, shard_id = mapping
if weight_name not in chunk_name:
continue
# Anyway, this is an expert weight and should not be
# attempted to load as other weights later
is_expert_weight = True
# Do not modify `name` since the loop may continue here
# Instead, create a new variable
name_mapped = chunk_name.replace(weight_name, param_name)
if is_pp_missing_parameter(name_mapped, self):
continue
param = params_dict[name_mapped]
# We should ask the weight loader to return success or
# not here since otherwise we may skip experts with
# other available replicas.
weight_loader = typing.cast(Callable[..., bool], param.weight_loader)
success = weight_loader(
param,
weight_to_load,
name_mapped,
shard_id=shard_id,
expert_id=expert_id,
return_success=True,
)
if success:
if not is_fusion_moe_shared_experts_layer:
name = name_mapped
else:
loaded_params.add(name_mapped)
break
else:
if is_expert_weight:
# We've checked that this is an expert weight
# However it's not mapped locally to this rank
# So we simply skip it
continue
# Skip loading extra bias for GPTQ models.
if name.endswith(".bias") and name not in params_dict:
continue
# Remapping the name of FP8 kv-scale.
name = maybe_remap_kv_scale_name(name, params_dict)
if name is None:
continue
if is_pp_missing_parameter(name, self):
continue
param = params_dict[name]
weight_loader = getattr(param, "weight_loader", default_weight_loader)
weight_loader(param, loaded_weight)
if not is_fusion_moe_shared_experts_layer:
loaded_params.add(name)
return loaded_params