完整部署快照

vllm/lora/ops/bgmv_expand.py

@@ -0,0 +1,168 @@
+"""
+Based on:
+Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023). 
+Punica: Multi-Tenant LoRA Serving. 
+https://arxiv.org/abs/2310.18547
+"""
+
+import torch
+import triton
+import triton.language as tl
+
+from .utils import get_lora_op_configs
+
+
+@triton.jit
+def _bgmv_expand_kernel(
+    input_ptr,
+    lora_ptr,
+    out_ptr,
+    N,
+    K,
+    lora_indices,
+    xm_stride,
+    xk_stride,
+    l0_stride,
+    lora_k_stride,
+    lora_n_stride,
+    cm_stride,
+    cn_stride,
+    BLOCK_N: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    SPLIT_N: tl.constexpr,
+    EVEN_K: tl.constexpr,
+    ADD_INPUTS: tl.constexpr,
+    CAST_TYPE: tl.constexpr,
+):
+    """
+    GroupGEMV, additionally, introducing SPLIT_N can improve large hidden_size's
+    performance
+    """
+    pid_sn = tl.program_id(axis=0)
+    cur_batch = tl.program_id(axis=1)
+    lora_index = tl.load(lora_indices + cur_batch)
+    if lora_index == -1:
+        return
+    offset_k = tl.arange(0, BLOCK_K)
+    offset_n = tl.arange(0, BLOCK_N)
+    if EVEN_K:
+        tiled_a = tl.load(input_ptr + cur_batch * xm_stride +
+                          offset_k * xk_stride, )  # [BLOCK_K]
+    else:
+        tiled_a = tl.load(
+            input_ptr + cur_batch * xm_stride + offset_k * xk_stride,
+            mask=offset_k < K,
+            other=0,
+        )  # [BLOCK_K]
+    # N must be divisible by SPLIT_N
+    split_n_length = tl.cdiv(N, SPLIT_N)
+    if CAST_TYPE:
+        tiled_a = tiled_a.to(lora_ptr.dtype.element_ty)
+    # sliding  to  next row-block
+    b_ptr = (lora_ptr + l0_stride * lora_index +
+             pid_sn * split_n_length * lora_k_stride)
+    c_ptr = out_ptr + cur_batch * cm_stride + pid_sn * split_n_length
+    for n in range(0, split_n_length, BLOCK_N):
+        current_n = n + offset_n
+        current_n_c = tl.max_contiguous(current_n, BLOCK_N)
+        b_ptr_mask = (current_n[:, None] < split_n_length) & (offset_k[None, :]
+                                                              < K)
+        c_mask = current_n < split_n_length
+        tiled_b = tl.load(
+            b_ptr + current_n_c[:, None] * lora_k_stride +
+            offset_k[None, :] * lora_n_stride,
+            mask=b_ptr_mask,
+            other=0.0,
+        )  # [BLOCK_N,BLOCK_K]
+        if ADD_INPUTS:
+            tiled_out = tl.load(c_ptr + current_n * cn_stride, mask=c_mask)
+            accumulator = tl.sum(tiled_a * tiled_b, 1) + tiled_out
+        else:
+            accumulator = tl.sum(tiled_a * tiled_b, 1)
+
+        tl.store(c_ptr + current_n * cn_stride, accumulator, mask=c_mask)
+
+
+@torch.inference_mode()
+def _bgmv_expand(
+    inputs: torch.Tensor,
+    lora_b_weights: torch.Tensor,
+    output_tensor: torch.Tensor,
+    lora_indices_tensor: torch.Tensor,
+    add_inputs: bool = True,
+) -> None:
+    """
+    Args:
+        inputs (torch.Tensor): input tensor
+        lora_b_weights (torch.Tensor): lora'a weight
+        output_tensor (torch.Tensor): output tensor
+        lora_indices_tensor (torch.Tensor): (batch_size,). The LoRA index
+            corresponding to each batch, An index of -1 means no lora should be
+            applied.
+        batches (int): batch size
+        add_inputs (bool, optional):  Defaults to False, adds the final lora 
+            results to the output.
+    """
+    assert inputs.dtype in [torch.float16, torch.bfloat16, torch.float32]
+    assert lora_b_weights.dtype in [
+        torch.float16,
+        torch.bfloat16,
+    ]
+    assert inputs.size(1) == lora_b_weights.size(-1)
+
+    assert inputs.is_contiguous()
+    assert output_tensor.is_contiguous()
+
+    if lora_b_weights.ndim == 4:  # shape:(lora_num,1,size,rank)
+        assert lora_b_weights.size(1) == 1
+        lora_b_weights = lora_b_weights.squeeze(dim=1)
+    else:
+        assert lora_b_weights.ndim == 3  # shape:(lora_num,size,rank)
+    assert lora_b_weights.is_contiguous()
+
+    # TODO tuning this config
+    N, K = lora_b_weights.shape[-2:]  # K= rank,N=hidden_size
+    BLOCK_K = triton.next_power_of_2(K)
+    EVEN_K = K % BLOCK_K == 0
+    ADD_INPUTS = add_inputs
+    CAST_TYPE = False
+    if inputs.dtype == torch.float32 and lora_b_weights.dtype in [
+            torch.float16,
+            torch.bfloat16,
+    ]:
+        CAST_TYPE = True
+    batches = lora_indices_tensor.size(0)
+    config = get_lora_op_configs("expand", batches, N)
+    grid = lambda META: (
+        META["SPLIT_N"],
+        batches,
+    )
+    _bgmv_expand_kernel[grid](
+        inputs,
+        lora_b_weights,
+        output_tensor,
+        N,
+        K,
+        lora_indices_tensor,
+        inputs.stride(0),
+        inputs.stride(1),
+        lora_b_weights.stride(0),
+        lora_b_weights.stride(1),
+        lora_b_weights.stride(2),
+        output_tensor.stride(0),
+        output_tensor.stride(1),
+        BLOCK_K=BLOCK_K,
+        EVEN_K=EVEN_K,
+        ADD_INPUTS=ADD_INPUTS,
+        CAST_TYPE=CAST_TYPE,
+        **config,
+    )
+    return
+
+
+try:
+    bgmv_expand = torch.library.custom_op("lora::bgmv_expand",
+                                          _bgmv_expand,
+                                          mutates_args=["output_tensor"])
+except AttributeError:
+    bgmv_expand = _bgmv_expand

vllm/lora/ops/bgmv_expand_slice.py

@@ -0,0 +1,181 @@
+"""
+Based on:
+Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023). 
+Punica: Multi-Tenant LoRA Serving. 
+https://arxiv.org/abs/2310.18547
+"""
+
+import torch
+import triton
+import triton.language as tl
+
+from .utils import get_lora_op_configs
+
+
+@triton.jit
+def _bgmv_expand_slice_kernel(
+    input_ptr,
+    lora_ptr,
+    out_ptr,
+    N,
+    K,
+    lora_indices,
+    xm_stride,
+    xk_stride,
+    l0_stride,
+    lora_k_stride,
+    lora_n_stride,
+    cm_stride,
+    cn_stride,
+    slice_offset,
+    BLOCK_N: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    SPLIT_N: tl.constexpr,
+    EVEN_K: tl.constexpr,
+    ADD_INPUTS: tl.constexpr,
+    CAST_TYPE: tl.constexpr,
+):
+    """
+    GroupGEMV, additionally, introducing SPLIT_N can improve large hidden_size's
+    performance
+    """
+    pid_sn = tl.program_id(axis=0)
+    cur_batch = tl.program_id(axis=1)
+    lora_index = tl.load(lora_indices + cur_batch)
+    if lora_index == -1:
+        return
+    offset_k = tl.arange(0, BLOCK_K)
+    offset_n = tl.arange(0, BLOCK_N)
+    if EVEN_K:
+        tiled_a = tl.load(input_ptr + cur_batch * xm_stride +
+                          offset_k * xk_stride, )  # [BLOCK_K]
+    else:
+        tiled_a = tl.load(
+            input_ptr + cur_batch * xm_stride + offset_k * xk_stride,
+            mask=offset_k < K,
+            other=0,
+        )  # [BLOCK_K]
+    # N must be divisible by SPLIT_N
+    split_n_length = tl.cdiv(N, SPLIT_N)
+    if CAST_TYPE:
+        tiled_a = tiled_a.to(lora_ptr.dtype.element_ty)
+    # sliding  to  next row-block
+    b_ptr = (lora_ptr + l0_stride * lora_index +
+             pid_sn * split_n_length * lora_k_stride)
+    c_ptr = (out_ptr + cur_batch * cm_stride + pid_sn * split_n_length +
+             slice_offset * cn_stride)
+
+    for n in range(0, split_n_length, BLOCK_N):
+        current_n = n + offset_n
+        b_ptr_mask = (current_n[:, None] < split_n_length) & (offset_k[None, :]
+                                                              < K)
+        c_mask = current_n < split_n_length
+        tiled_b = tl.load(
+            b_ptr + current_n[:, None] * lora_k_stride +
+            offset_k[None, :] * lora_n_stride,
+            mask=b_ptr_mask,
+            other=0.0,
+        )  # [BLOCK_N,BLOCK_K]
+
+        if ADD_INPUTS:
+            tiled_out = tl.load(c_ptr + current_n * cn_stride, mask=c_mask)
+            accumulator = tl.sum(tiled_a * tiled_b, 1) + tiled_out
+        else:
+            accumulator = tl.sum(tiled_a * tiled_b, 1)
+
+        tl.store(c_ptr + current_n * cn_stride, accumulator, mask=c_mask)
+
+
+@torch.inference_mode()
+def _bgmv_expand_slice(
+    inputs: torch.Tensor,
+    lora_b_weights: torch.Tensor,
+    output_tensor: torch.Tensor,
+    lora_indices_tensor: torch.Tensor,
+    slice_offset: int,
+    slice_size: int,
+    add_inputs: bool = True,
+) -> None:
+    """
+    Args:
+        inputs (torch.Tensor): input tensor
+        lora_b_weights (torch.Tensor): lora'b weight
+        output_tensor (torch.Tensor): output tensor
+        lora_indices_tensor (torch.Tensor): (batch_size,). The LoRA index
+            corresponding to each batch, An index of -1 means no lora should be
+            applied.
+        slice_offset (int): output_tensor's offset
+        slice_size (int): current output_tensor's size
+        batches (int): batch size
+        add_inputs (bool, optional): Defaults to False.
+    """
+    assert inputs.dtype in [torch.float16, torch.bfloat16, torch.float32]
+    assert lora_b_weights.dtype in [
+        torch.float16,
+        torch.bfloat16,
+    ]
+    assert inputs.size(1) == lora_b_weights.size(-1)
+
+    assert slice_size == lora_b_weights.size(-2)
+    assert inputs.is_contiguous()
+    assert output_tensor.is_contiguous()
+
+    if lora_b_weights.ndim == 4:  # shape:(lora_num,1,size,rank)
+        assert lora_b_weights.size(1) == 1
+        lora_b_weights = lora_b_weights.squeeze(dim=1)
+    else:
+        assert lora_b_weights.ndim == 3  # shape:(lora_num,size,rank)
+
+    assert lora_b_weights.is_contiguous()
+
+    # TODO tuning this config
+
+    N, K = lora_b_weights.shape[-2:]  # K= rank,N=hidden_size
+    BLOCK_K = triton.next_power_of_2(K)
+    EVEN_K = K % BLOCK_K == 0
+    ADD_INPUTS = add_inputs
+    CAST_TYPE = False
+    if inputs.dtype == torch.float32 and lora_b_weights.dtype in [
+            torch.float16,
+            torch.bfloat16,
+    ]:
+        CAST_TYPE = True
+
+    batches = lora_indices_tensor.size(0)
+
+    config = get_lora_op_configs("expand", batches, N)
+
+    grid = lambda META: (
+        META["SPLIT_N"],
+        batches,
+    )
+    _bgmv_expand_slice_kernel[grid](
+        inputs,
+        lora_b_weights,
+        output_tensor,
+        N,
+        K,
+        lora_indices_tensor,
+        inputs.stride(0),
+        inputs.stride(1),
+        lora_b_weights.stride(0),
+        lora_b_weights.stride(1),
+        lora_b_weights.stride(2),
+        output_tensor.stride(0),
+        output_tensor.stride(1),
+        slice_offset,
+        BLOCK_K=BLOCK_K,
+        EVEN_K=EVEN_K,
+        ADD_INPUTS=ADD_INPUTS,
+        CAST_TYPE=CAST_TYPE,
+        **config,
+    )
+    return
+
+
+try:
+    bgmv_expand_slice = torch.library.custom_op("lora::bgmv_expand_slice",
+                                                _bgmv_expand_slice,
+                                                mutates_args=["output_tensor"])
+except AttributeError:
+    bgmv_expand_slice = _bgmv_expand_slice

vllm/lora/ops/bgmv_shrink.py

@@ -0,0 +1,150 @@
+"""
+Based on:
+Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023). 
+Punica: Multi-Tenant LoRA Serving. 
+https://arxiv.org/abs/2310.18547
+"""
+
+import torch
+import triton
+import triton.language as tl
+
+from .utils import get_lora_op_configs
+
+
+@triton.jit
+def _bgmv_shrink_kernel(
+    input_ptr,
+    lora_ptr,
+    out_ptr,
+    N,
+    K,
+    lora_indices,
+    scaling,
+    xm_stride,
+    xk_stride,
+    l0_stride,
+    lora_k_stride,
+    lora_n_stride,
+    cm_stride,
+    cn_stride,
+    BLOCK_N: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    SPLIT_K: tl.constexpr,
+):
+    """
+    GroupGEMV, additionally, introducing SPLIT-K can improve large hidden_size's
+    performance
+    """
+    pid_sk = tl.program_id(axis=0)
+    cur_batch = tl.program_id(axis=1)
+    lora_index = tl.load(lora_indices + cur_batch)
+    if lora_index == -1:
+        return
+
+    offset_n = tl.arange(0, BLOCK_N)
+    offset_k = tl.arange(0, BLOCK_K) + pid_sk * BLOCK_K
+    a_ptr = input_ptr + cur_batch * xm_stride
+    b_ptr = lora_ptr + l0_stride * lora_index
+    accumulator = tl.zeros((BLOCK_N, ), dtype=tl.float32)
+    for k in range(0, K, BLOCK_K * SPLIT_K):
+        current_k = k + offset_k
+        current_k_c = tl.max_contiguous(current_k, BLOCK_K)
+        tiled_a = tl.load(
+            a_ptr + current_k_c,
+            mask=current_k < K,
+            other=0.0,
+        )  # [BLOCK_K]
+        b_ptr_mask = (offset_n[:, None] < N) & (current_k[None, :] < K)
+
+        tiled_b = tl.load(
+            b_ptr + offset_n[:, None] * lora_k_stride +
+            current_k[None, :] * lora_n_stride,
+            mask=b_ptr_mask,
+            other=0.0,
+        )  # [BLOCK_N,BLOCK_K]
+
+        accumulator += tl.sum(tiled_a * tiled_b, 1)
+    accumulator *= scaling
+    offset_cn = tl.arange(0, BLOCK_N)
+    c_ptr = out_ptr + cur_batch * cm_stride + offset_cn * cn_stride
+    c_mask = offset_cn < N
+    if SPLIT_K == 1:
+        tl.store(c_ptr, accumulator, mask=c_mask)
+    else:
+        tl.atomic_add(c_ptr, accumulator, mask=c_mask)
+
+
+@torch.inference_mode()
+def _bgmv_shrink(
+    inputs: torch.Tensor,
+    lora_a_weights: torch.Tensor,
+    output_tensor: torch.Tensor,
+    lora_indices_tensor: torch.Tensor,
+    scaling: float = 1.0,
+) -> None:
+    """
+    Args:
+        inputs (torch.Tensor): input tensor
+        lora_a_weights (torch.Tensor): lora'a weight
+        output_tensor (torch.Tensor): output tensor
+        lora_indices_tensor (torch.Tensor): (batch_size,). The LoRA index
+            corresponding to each batch. An index of -1 means no lora should be
+            applied.
+        batches (int): batch size
+        scaling (float):  Scaling factor.
+    """
+    assert inputs.dtype == lora_a_weights.dtype
+    assert inputs.dtype in [torch.float16, torch.bfloat16]
+    assert lora_a_weights.dtype in [
+        torch.float16,
+        torch.bfloat16,
+    ]
+    assert inputs.size(1) == lora_a_weights.size(-1)
+    assert inputs.is_contiguous()
+
+    if lora_a_weights.ndim == 4:  # shape:(lora_num,1,rank, size)
+        assert lora_a_weights.size(1) == 1
+        lora_a_weights = lora_a_weights.squeeze(dim=1)
+    else:
+        assert lora_a_weights.ndim == 3  # shape:(lora_num,rank, size)
+    assert lora_a_weights.is_contiguous()
+    assert output_tensor.is_contiguous()
+    # TODO tuning this config
+    batches = lora_indices_tensor.size(0)
+    N, K = lora_a_weights.shape[-2:]  # K=hidden_size,N=rank
+    BLOCK_N = triton.next_power_of_2(N)
+    # First try to load optimal config from the file
+    config = get_lora_op_configs("bgmv_shrink", batches, K)
+
+    grid = lambda META: (
+        META["SPLIT_K"],
+        batches,
+    )
+    _bgmv_shrink_kernel[grid](
+        inputs,
+        lora_a_weights,
+        output_tensor,
+        N,
+        K,
+        lora_indices_tensor,
+        scaling,
+        inputs.stride(0),
+        inputs.stride(1),
+        lora_a_weights.stride(0),
+        lora_a_weights.stride(1),
+        lora_a_weights.stride(2),
+        output_tensor.stride(0),
+        output_tensor.stride(1),
+        BLOCK_N=BLOCK_N,
+        **config,
+    )
+    return
+
+
+try:
+    bgmv_shrink = torch.library.custom_op("lora::bgmv_shrink",
+                                          _bgmv_shrink,
+                                          mutates_args=["output_tensor"])
+except AttributeError:
+    bgmv_shrink = _bgmv_shrink

vllm/lora/ops/sgmv_expand.py

@@ -0,0 +1,204 @@
+"""
+Based on:
+Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023). 
+Punica: Multi-Tenant LoRA Serving. 
+https://arxiv.org/abs/2310.18547
+"""
+
+import torch
+import triton
+import triton.language as tl
+
+from vllm.triton_utils import libentry
+
+
+@libentry()
+@triton.jit
+def _sgmv_expand_kernel(
+    input_ptr,
+    lora_ptr,
+    out_ptr,
+    N,
+    K,
+    b_seq_start_loc,
+    seq_lens,
+    lora_indices,
+    xm_stride,
+    xk_stride,  # 1
+    l0_stride,  # hidden_size*max_rank
+    lora_k_stride,
+    lora_n_stride,
+    cm_stride,
+    cn_stride,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    EVEN_K: tl.constexpr,
+    ADD_INPUTS: tl.constexpr,
+    CAST_TYPE: tl.constexpr,
+):
+    """
+    The sgmv's expand triton kernel is based on GroupGEMM.
+    """
+    pid = tl.program_id(axis=0)
+    cur_batch = tl.program_id(axis=1)
+    cta_n_num = tl.cdiv(N, BLOCK_N)
+    pid_m = pid // cta_n_num
+    pid_n = pid % cta_n_num
+    M = tl.load(seq_lens + cur_batch)
+    if pid_m * BLOCK_M > M:
+        return
+    lora_index = tl.load(lora_indices + cur_batch)
+    if lora_index == -1:
+        return
+    cur_seq_start = tl.load(b_seq_start_loc + cur_batch)
+    offset_m = tl.arange(0, BLOCK_M) + pid_m * BLOCK_M
+    offset_n = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N
+    offset_k = tl.arange(0, BLOCK_K)
+    ram = tl.max_contiguous(tl.multiple_of(offset_m % M, BLOCK_M), BLOCK_M)
+    rbn = tl.max_contiguous(tl.multiple_of(offset_n % N, BLOCK_N), BLOCK_N)
+
+    a_ptr = (input_ptr + cur_seq_start * xm_stride + ram[:, None] * xm_stride +
+             offset_k[None, :] * xk_stride, )
+    b_ptr = (lora_ptr + l0_stride * lora_index +
+             offset_k[:, None] * lora_n_stride + rbn[None, :] * lora_k_stride)
+    accumulator = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
+    for k in range(tl.cdiv(K, BLOCK_K)):
+        if EVEN_K:
+            tiled_a = tl.load(a_ptr)
+            tiled_b = tl.load(b_ptr)
+        else:
+            tiled_a = tl.load(a_ptr,
+                              mask=offset_k[None, :] < K - k * BLOCK_K,
+                              other=0)
+            tiled_b = tl.load(b_ptr,
+                              mask=offset_k[:, None] < K - k * BLOCK_K,
+                              other=0)
+        if CAST_TYPE:
+            tiled_a = tiled_a.to(lora_ptr.dtype.element_ty)
+        accumulator += tl.dot(
+            tiled_a,
+            tiled_b,
+        )
+        a_ptr += BLOCK_K * xk_stride
+        b_ptr += BLOCK_K * lora_n_stride
+    tiled_c = accumulator.to(lora_ptr.dtype.element_ty)
+    offset_cm = cur_seq_start + tl.arange(0, BLOCK_M) + pid_m * BLOCK_M
+    offset_cn = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N
+    c_ptr = (out_ptr + offset_cm[:, None] * cm_stride +
+             offset_cn[None, :] * cn_stride)
+    M = tl.load(seq_lens + cur_batch)
+    c_mask = (offset_cm[:, None] <
+              (cur_seq_start + M)) & (offset_cn[None, :] < N)
+    if ADD_INPUTS:
+        tiled_out = tl.load(c_ptr, mask=c_mask)
+        tiled_c += tiled_out
+    tl.store(c_ptr, tiled_c, mask=c_mask)
+
+
+@torch.inference_mode()
+def _sgmv_expand(
+    inputs: torch.Tensor,
+    lora_b_weights: torch.Tensor,
+    output_tensor: torch.Tensor,
+    b_seq_start_loc: torch.Tensor,
+    seq_len_tensor: torch.Tensor,
+    lora_indices_tensor: torch.Tensor,
+    batches: int,
+    max_seq_length: int,
+    token_nums: int,
+    add_inputs: bool = False,
+) -> None:
+    """
+    Args:
+        inputs (torch.Tensor): input tensor
+        lora_b_weights (torch.Tensor): lora'a weight
+        output_tensor (torch.Tensor): output tensor
+        b_seq_start_loc (torch.Tensor): (batch_size,). The cumulative
+            sequence lengths of the sequences in the batch, used to index
+            into sequence. E.g., if the sequence length is [4, 6], it is
+            [0, 4, 10].
+        seq_len_tensor (torch.Tensor): (batch_size,). Record the sequence
+            length of the sequences in the batch.
+        lora_indices_tensor (torch.Tensor): (batch_size,). The LoRA index
+            corresponding to each batch. An index of -1 means no lora should be
+            applied.
+        batches (int): batch size
+        max_seq_length (int): The max sequence lengths of the sequences in the 
+            batch.
+        token_nums (int): The token numbers in the batch. Used to verify if the 
+            token numbers in the inputs matches the one in the metadata.
+        add_inputs (bool, optional): Defaults to False, adds the final lora 
+            results to the output.
+    """
+
+    assert inputs.dtype in [torch.float16, torch.bfloat16, torch.float32]
+    assert lora_b_weights.dtype in [
+        torch.float16,
+        torch.bfloat16,
+    ]
+    assert inputs.size(0) == token_nums
+    assert inputs.size(1) == lora_b_weights.size(-1)
+    assert b_seq_start_loc.size(0) == batches
+    assert lora_indices_tensor.size(0) == batches
+    assert inputs.is_contiguous()
+    assert output_tensor.is_contiguous()
+
+    if lora_b_weights.ndim == 4:  # shape:(lora_num,1,size,rank)
+        assert lora_b_weights.size(1) == 1
+        lora_b_weights = lora_b_weights.squeeze(dim=1)
+    else:
+        assert lora_b_weights.ndim == 3  # shape:(lora_num,size,rank)
+
+    assert lora_b_weights.is_contiguous()
+
+    # TODO tuning this config
+
+    N, K = lora_b_weights.shape[-2:]  # K= rank,N=hidden_size
+    BLOCK_M = 32
+    BLOCK_N = 32
+    BLOCK_K = 16
+    EVEN_K = K % BLOCK_K == 0
+    ADD_INPUTS = add_inputs
+    CAST_TYPE = False
+    if inputs.dtype == torch.float32 and lora_b_weights.dtype in [
+            torch.float16,
+            torch.bfloat16,
+    ]:
+        CAST_TYPE = True
+    grid = (
+        triton.cdiv(max_seq_length, BLOCK_M) * triton.cdiv(N, BLOCK_N),
+        batches,
+    )
+    _sgmv_expand_kernel[grid](
+        inputs,
+        lora_b_weights,
+        output_tensor,
+        N,
+        K,
+        b_seq_start_loc,
+        seq_len_tensor,
+        lora_indices_tensor,
+        inputs.stride(0),
+        inputs.stride(1),
+        lora_b_weights.stride(0),
+        lora_b_weights.stride(1),
+        lora_b_weights.stride(2),
+        output_tensor.stride(0),
+        output_tensor.stride(1),
+        BLOCK_M,
+        BLOCK_N,
+        BLOCK_K,
+        EVEN_K,
+        ADD_INPUTS,
+        CAST_TYPE,
+    )
+    return
+
+
+try:
+    sgmv_expand = torch.library.custom_op("lora::sgmv_expand",
+                                          _sgmv_expand,
+                                          mutates_args=["output_tensor"])
+except AttributeError:
+    sgmv_expand = _sgmv_expand

vllm/lora/ops/sgmv_expand_slice.py

@@ -0,0 +1,217 @@
+"""
+Based on:
+Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023). 
+Punica: Multi-Tenant LoRA Serving. 
+https://arxiv.org/abs/2310.18547
+"""
+
+import torch
+import triton
+import triton.language as tl
+
+from vllm.triton_utils import libentry
+
+
+@libentry()
+@triton.jit
+def _sgmv_expand_slice_kernel(
+    input_ptr,
+    lora_ptr,
+    out_ptr,
+    N,
+    K,
+    b_seq_start_loc,
+    seq_lens,
+    lora_indices,
+    xm_stride,
+    xk_stride,  # 1
+    l0_stride,  # hidden_size*max_rank
+    lora_k_stride,
+    lora_n_stride,
+    cm_stride,
+    cn_stride,
+    slice_offset,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    EVEN_K: tl.constexpr,
+    ADD_INPUTS: tl.constexpr,
+    CAST_TYPE: tl.constexpr,
+):
+    """
+
+    Similar to the 'sgmv_expand' operator, but with an added parameter 
+    'slice_offset'. The reason for not reusing the 'sgmv_expand' operator 
+    might be that in the future, we could implement a fusion operator to 
+    achieve the current functionality instead of having to call it multiple 
+    times.
+    """
+    pid = tl.program_id(axis=0)
+    cur_batch = tl.program_id(axis=1)
+    cta_n_num = tl.cdiv(N, BLOCK_N)
+    pid_m = pid // cta_n_num
+    pid_n = pid % cta_n_num
+    M = tl.load(seq_lens + cur_batch)
+    if pid_m * BLOCK_M > M:
+        return
+    lora_index = tl.load(lora_indices + cur_batch)
+    if lora_index == -1:
+        return
+    cur_seq_start = tl.load(b_seq_start_loc + cur_batch)
+    offset_m = tl.arange(0, BLOCK_M) + pid_m * BLOCK_M
+    offset_n = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N
+    offset_k = tl.arange(0, BLOCK_K)
+    ram = tl.max_contiguous(tl.multiple_of(offset_m % M, BLOCK_M), BLOCK_M)
+    rbn = tl.max_contiguous(tl.multiple_of(offset_n % N, BLOCK_N), BLOCK_N)
+
+    a_ptr = (input_ptr + cur_seq_start * xm_stride + ram[:, None] * xm_stride +
+             offset_k[None, :] * xk_stride, )
+    b_ptr = (lora_ptr + l0_stride * lora_index +
+             offset_k[:, None] * lora_n_stride + rbn[None, :] * lora_k_stride)
+    accumulator = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
+    for k in range(tl.cdiv(K, BLOCK_K)):
+        if EVEN_K:
+            tiled_a = tl.load(a_ptr)
+            tiled_b = tl.load(b_ptr)
+        else:
+            tiled_a = tl.load(a_ptr,
+                              mask=offset_k[None, :] < K - k * BLOCK_K,
+                              other=0)
+            tiled_b = tl.load(b_ptr,
+                              mask=offset_k[:, None] < K - k * BLOCK_K,
+                              other=0)
+        if CAST_TYPE:
+            tiled_a = tiled_a.to(lora_ptr.dtype.element_ty)
+        accumulator += tl.dot(
+            tiled_a,
+            tiled_b,
+        )
+        a_ptr += BLOCK_K * xk_stride
+        b_ptr += BLOCK_K * lora_n_stride
+    tiled_c = accumulator.to(lora_ptr.dtype.element_ty)
+    offset_cm = cur_seq_start + tl.arange(0, BLOCK_M) + pid_m * BLOCK_M
+    offset_cn = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N + slice_offset
+    c_ptr = (out_ptr + offset_cm[:, None] * cm_stride +
+             offset_cn[None, :] * cn_stride)
+    M = tl.load(seq_lens + cur_batch)
+    c_mask = (offset_cm[:, None] < (cur_seq_start + M)) & (offset_cn[None, :] <
+                                                           (slice_offset + N))
+    if ADD_INPUTS:
+        tiled_out = tl.load(c_ptr, mask=c_mask)
+        tiled_c += tiled_out
+    tl.store(c_ptr, tiled_c, mask=c_mask)
+
+
+@torch.inference_mode()
+def _sgmv_expand_slice(
+    inputs: torch.Tensor,
+    lora_b_weights: torch.Tensor,
+    output_tensor: torch.Tensor,
+    b_seq_start_loc: torch.Tensor,
+    seq_len_tensor: torch.Tensor,
+    lora_indices_tensor: torch.Tensor,
+    batches: int,
+    max_seq_length: int,
+    token_nums: int,
+    slice_offset: int,
+    slice_size: int,
+    add_inputs: bool = False,
+) -> None:
+    """_summary_
+
+    Args:
+        inputs (torch.Tensor): input tensor
+        lora_b_weights (torch.Tensor): lora'a weight
+        output_tensor (torch.Tensor): output tensor
+        b_seq_start_loc (torch.Tensor): (batch_size,). The cumulative
+            sequence lengths of the sequences in the batch, used to index
+            into sequence. E.g., if the sequence length is [4, 6], it is
+            [0, 4, 10].
+        seq_len_tensor (torch.Tensor): (batch_size,). Record the sequence
+            length of the sequences in the batch
+        lora_indices_tensor (torch.Tensor): (batch_size,). The LoRA index
+            corresponding to each batch. An index of -1 means no lora should be
+            applied.
+        batches (int): batch size
+        max_seq_length (int): The max sequence lengths of the sequences
+            in the batch
+        token_nums (int): The token numbers in the batch. Used to verify if the 
+            token numbers in the inputs matches the one in the metadata.
+        slice_offset (int): output_tensor's offset
+        slice_size (int): current output_tensor's size
+        add_inputs (bool, optional): Defaults to False, adds the final lora 
+            results to the output.
+    """
+
+    assert inputs.dtype in [torch.float16, torch.bfloat16, torch.float32]
+    assert lora_b_weights.dtype in [
+        torch.float16,
+        torch.bfloat16,
+    ]
+    assert inputs.size(0) == token_nums
+    assert inputs.size(1) == lora_b_weights.size(-1)
+    assert b_seq_start_loc.size(0) == batches
+    assert lora_indices_tensor.size(0) == batches
+    assert slice_size == lora_b_weights.size(-2)
+    assert inputs.is_contiguous()
+    assert output_tensor.is_contiguous()
+
+    if lora_b_weights.ndim == 4:  # shape:(lora_num,1,size,rank)
+        assert lora_b_weights.size(1) == 1
+        lora_b_weights = lora_b_weights.squeeze(dim=1)
+    else:
+        assert lora_b_weights.ndim == 3  # shape:(lora_num,size,rank)
+
+    assert lora_b_weights.is_contiguous()
+
+    # TODO tuning this config
+    N, K = lora_b_weights.shape[-2:]  # K= rank,N=hidden_size
+
+    BLOCK_M = 32
+    BLOCK_N = 32
+    BLOCK_K = 16
+    EVEN_K = K % BLOCK_K == 0
+    ADD_INPUTS = add_inputs
+    CAST_TYPE = False
+    if inputs.dtype == torch.float32 and lora_b_weights.dtype in [
+            torch.float16,
+            torch.bfloat16,
+    ]:
+        CAST_TYPE = True
+    grid = (
+        triton.cdiv(max_seq_length, BLOCK_M) * triton.cdiv(N, BLOCK_N),
+        batches,
+    )
+    _sgmv_expand_slice_kernel[grid](
+        inputs,
+        lora_b_weights,
+        output_tensor,
+        N,
+        K,
+        b_seq_start_loc,
+        seq_len_tensor,
+        lora_indices_tensor,
+        inputs.stride(0),
+        inputs.stride(1),
+        lora_b_weights.stride(0),
+        lora_b_weights.stride(1),
+        lora_b_weights.stride(2),
+        output_tensor.stride(0),
+        output_tensor.stride(1),
+        slice_offset,
+        BLOCK_M,
+        BLOCK_N,
+        BLOCK_K,
+        EVEN_K,
+        ADD_INPUTS,
+        CAST_TYPE,
+    )
+    return
+
+
+try:
+    sgmv_expand_slice = torch.library.custom_op("lora::sgmv_expand_slice",
+                                                _sgmv_expand_slice,
+                                                mutates_args=["output_tensor"])
+except AttributeError:
+    sgmv_expand_slice = _sgmv_expand_slice

vllm/lora/ops/sgmv_shrink.py

@@ -0,0 +1,201 @@
+"""
+Based on:
+Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023). 
+Punica: Multi-Tenant LoRA Serving. 
+https://arxiv.org/abs/2310.18547
+"""
+
+import torch
+import triton
+import triton.language as tl
+
+from vllm.triton_utils import libentry
+
+
+@libentry()
+@triton.jit
+def _sgmv_shrink_kernel(
+    input_ptr,
+    lora_ptr,
+    out_ptr,
+    N,
+    K,
+    b_seq_start_loc,
+    seq_lens,
+    lora_indices,
+    scaling,
+    xm_stride,  # hidden_size
+    xk_stride,  # 1
+    l0_stride,  # hidden_size*max_rank
+    lora_k_stride,
+    lora_n_stride,
+    cm_stride,
+    cn_stride,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    EVEN_K: tl.constexpr,
+    SPLIT_K: tl.constexpr,
+):
+    """
+    The sgmv's shrink triton kernel is based on GroupGEMM+SPLIT-K.
+    The GEMM of Multi-LoRA can be considered as GroupGEMM. Additionally,
+    introducing SPLIT-K can improve performance
+    """
+    pid = tl.program_id(axis=0)
+    pid_sk = tl.program_id(axis=1)
+    cur_batch = tl.program_id(axis=2)
+    cta_n_num = tl.cdiv(N, BLOCK_N)
+    pid_m = pid // cta_n_num
+    pid_n = pid % cta_n_num
+
+    M = tl.load(seq_lens + cur_batch)
+    if pid_m * BLOCK_M > M:
+        return
+    lora_index = tl.load(lora_indices + cur_batch)
+    if lora_index == -1:
+        return
+    cur_seq_start = tl.load(b_seq_start_loc + cur_batch)
+    offset_m = tl.arange(0, BLOCK_M) + pid_m * BLOCK_M
+    offset_n = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N
+    offset_k = pid_sk * BLOCK_K + tl.arange(0, BLOCK_K)
+
+    ram = tl.max_contiguous(tl.multiple_of(offset_m % M, BLOCK_M), BLOCK_M)
+    rbn = tl.max_contiguous(tl.multiple_of(offset_n % N, BLOCK_N), BLOCK_N)
+
+    a_ptr = (input_ptr + cur_seq_start * xm_stride + ram[:, None] * xm_stride +
+             offset_k[None, :] * xk_stride)
+    b_ptr = (lora_ptr + l0_stride * lora_index + rbn[None, :] * lora_k_stride +
+             offset_k[:, None] * lora_n_stride)
+
+    accumulator = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
+    for k in range(0, tl.cdiv(K, BLOCK_K * SPLIT_K)):
+        if EVEN_K:
+            tiled_a = tl.load(a_ptr)
+            tiled_b = tl.load(b_ptr)
+        else:
+            k_remaining = K - k * (BLOCK_K * SPLIT_K)
+            tiled_a = tl.load(a_ptr,
+                              mask=offset_k[None, :] < k_remaining,
+                              other=0.0)
+            tiled_b = tl.load(b_ptr,
+                              mask=offset_k[:, None] < k_remaining,
+                              other=0.0)
+        accumulator += tl.dot(tiled_a, tiled_b)
+
+        a_ptr += BLOCK_K * SPLIT_K * xk_stride
+        b_ptr += BLOCK_K * SPLIT_K * lora_n_stride
+    offset_cm = cur_seq_start + tl.arange(0, BLOCK_M) + pid_m * BLOCK_M
+
+    offset_cn = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N
+    c_ptr = (out_ptr + offset_cm[:, None] * cm_stride +
+             offset_cn[None, :] * cn_stride)
+    c_mask = (offset_cm[:, None] <
+              (cur_seq_start + M)) & (offset_cn[None, :] < N)
+    accumulator *= scaling
+    # handles write-back with reduction-splitting
+    if SPLIT_K == 1:
+        tl.store(c_ptr, accumulator, mask=c_mask)
+    else:
+        tl.atomic_add(c_ptr, accumulator, mask=c_mask)
+
+
+@torch.inference_mode()
+def _sgmv_shrink(
+    inputs: torch.Tensor,
+    lora_a_weights: torch.Tensor,
+    output_tensor: torch.Tensor,
+    b_seq_start_loc: torch.Tensor,
+    seq_len_tensor: torch.Tensor,
+    lora_indices_tensor: torch.Tensor,
+    batches: int,
+    max_seq_length: int,
+    token_nums: int,
+    scaling: float,
+) -> None:
+    """
+
+    Args:
+        inputs (torch.Tensor): input tensor
+        lora_a_weights (torch.Tensor): lora'a weight
+        output_tensor (torch.Tensor): output tensor
+        b_seq_start_loc (torch.Tensor): (batch_size,). The cumulative
+            sequence lengths of the sequences in the batch, used to index
+            into sequence. E.g., if the sequence length is [4, 6], it is
+            [0, 4].
+        seq_len_tensor (torch.Tensor): (batch_size,). Record the sequence
+            length of the sequences in the batch.
+        lora_indices_tensor (torch.Tensor): (batch_size,). The LoRA index
+            corresponding to each batch. An index of -1 means no lora should be
+            applied.
+        batches (int): batch size
+        max_seq_length (int): The max sequence lengths of the sequences in the 
+            batch.
+        token_nums (int): The token numbers in the batch. Used to verify if the 
+            token numbers in the inputs matches the one in the metadata.
+        scaling (float): Scaling factor.
+    """
+    assert inputs.dtype == lora_a_weights.dtype
+    assert inputs.dtype in [torch.float16, torch.bfloat16]
+    assert lora_a_weights.dtype in [
+        torch.float16,
+        torch.bfloat16,
+    ]
+    assert inputs.size(0) == token_nums
+    assert inputs.size(1) == lora_a_weights.size(-1)
+    assert b_seq_start_loc.size(0) == batches
+    assert lora_indices_tensor.size(0) == batches
+    assert inputs.is_contiguous()
+
+    if lora_a_weights.ndim == 4:  # shape:(lora_num,1,rank, size)
+        assert lora_a_weights.size(1) == 1
+        lora_a_weights = lora_a_weights.squeeze(dim=1)
+    else:
+        assert lora_a_weights.ndim == 3  # shape:(lora_num,rank, size)
+    assert lora_a_weights.is_contiguous()
+    assert output_tensor.is_contiguous()
+    # TODO tuning this config
+    N, K = lora_a_weights.shape[-2:]  # K=hidden_size,N=rank
+    BLOCK_M = 32
+    BLOCK_N = 16
+    BLOCK_K = 32
+    SPLIT_K = 8
+    EVEN_K = K % (BLOCK_K * SPLIT_K) == 0
+    grid = (
+        triton.cdiv(max_seq_length, BLOCK_M) * triton.cdiv(N, BLOCK_N),
+        SPLIT_K,
+        batches,
+    )
+
+    _sgmv_shrink_kernel[grid](
+        inputs,
+        lora_a_weights,
+        output_tensor,
+        N,
+        K,
+        b_seq_start_loc,
+        seq_len_tensor,
+        lora_indices_tensor,
+        scaling,
+        inputs.stride(0),
+        inputs.stride(1),
+        lora_a_weights.stride(0),
+        lora_a_weights.stride(1),
+        lora_a_weights.stride(2),
+        output_tensor.stride(0),
+        output_tensor.stride(1),
+        BLOCK_M,
+        BLOCK_N,
+        BLOCK_K,
+        EVEN_K,
+        SPLIT_K,
+    )
+    return
+
+
+try:
+    sgmv_shrink = torch.library.custom_op("lora::sgmv_shrink",
+                                          _sgmv_shrink,
+                                          mutates_args=["output_tensor"])
+except AttributeError:
+    sgmv_shrink = _sgmv_shrink

vllm/lora/ops/torch_ops/__init__.py

@@ -0,0 +1,15 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from vllm.lora.ops.torch_ops.lora_ops import bgmv_expand  # noqa: F401
+from vllm.lora.ops.torch_ops.lora_ops import (bgmv_expand_slice, bgmv_shrink,
+                                              sgmv_expand, sgmv_expand_slice,
+                                              sgmv_shrink)
+
+__all__ = [
+    "bgmv_expand",
+    "bgmv_expand_slice",
+    "bgmv_shrink",
+    "sgmv_expand",
+    "sgmv_expand_slice",
+    "sgmv_shrink",
+]

vllm/lora/ops/torch_ops/lora_ops.py

@@ -0,0 +1,115 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+
+
+def sgmv_expand(inputs: torch.Tensor,
+                lora_b_weights: torch.Tensor,
+                output_tensor: torch.Tensor,
+                b_seq_start_loc: torch.Tensor,
+                seq_len_tensor: torch.Tensor,
+                lora_indices_tensor: torch.Tensor,
+                batches: int,
+                max_seq_length: int,
+                token_nums: int,
+                add_inputs: bool = False):
+    exploded_indices = torch.repeat_interleave(lora_indices_tensor,
+                                               seq_len_tensor)
+
+    bgmv_expand(inputs, lora_b_weights, output_tensor, exploded_indices,
+                add_inputs)
+
+
+def bgmv_expand(inputs: torch.Tensor,
+                lora_b_weights: torch.Tensor,
+                output_tensor: torch.Tensor,
+                lora_indices_tensor: torch.Tensor,
+                add_inputs: bool = True):
+    selected_loras = lora_b_weights[lora_indices_tensor].to(
+        dtype=output_tensor.dtype)
+    if len(selected_loras.shape) == 4:
+        selected_loras = selected_loras.squeeze(dim=1)
+    inputs = inputs.to(dtype=output_tensor.dtype)
+    outputs = torch.einsum("bi, boi -> bo", inputs, selected_loras)
+
+    limit = output_tensor.shape[0]
+    if outputs.shape[0] == 1 and output_tensor.shape[0] != 1:
+        limit = 1
+
+    if add_inputs:
+        output_tensor[:, :outputs.shape[1]] += outputs[:limit, :]
+    else:
+        output_tensor[:, :outputs.shape[1]] = outputs[:limit, :]
+
+
+def sgmv_shrink(
+    inputs: torch.Tensor,
+    lora_a_weights: torch.Tensor,
+    output_tensor: torch.Tensor,
+    b_seq_start_loc: torch.Tensor,
+    seq_len_tensor: torch.Tensor,
+    lora_indices_tensor: torch.Tensor,
+    batches: int,
+    max_seq_length: int,
+    token_nums: int,
+    scaling: float,
+):
+    exploded_indices = torch.repeat_interleave(lora_indices_tensor,
+                                               seq_len_tensor)
+
+    bgmv_shrink(inputs, lora_a_weights, output_tensor, exploded_indices,
+                scaling)
+
+
+def bgmv_shrink(inputs: torch.Tensor,
+                lora_b_weights: torch.Tensor,
+                output_tensor: torch.Tensor,
+                lora_indices_tensor: torch.Tensor,
+                scaling: float = 1.0):
+    selected_loras = lora_b_weights[lora_indices_tensor].to(
+        dtype=output_tensor.dtype)
+    if len(selected_loras.shape) == 4:
+        selected_loras = selected_loras.squeeze(dim=1)
+    inputs = inputs.to(dtype=output_tensor.dtype)
+    outputs = torch.einsum("bi, boi -> bo", inputs, selected_loras)
+
+    output_tensor[:, :outputs.shape[1]] = scaling * outputs[:]
+
+
+def sgmv_expand_slice(inputs: torch.Tensor,
+                      lora_b_weights: torch.Tensor,
+                      output_tensor: torch.Tensor,
+                      b_seq_start_loc: torch.Tensor,
+                      seq_len_tensor: torch.Tensor,
+                      lora_indices_tensor: torch.Tensor,
+                      batches: int,
+                      max_seq_length: int,
+                      token_nums: int,
+                      slice_offset: int,
+                      slice_size: int,
+                      add_inputs: bool = False):
+    exploded_indices = torch.repeat_interleave(lora_indices_tensor,
+                                               seq_len_tensor)
+
+    bgmv_expand_slice(inputs, lora_b_weights, output_tensor, exploded_indices,
+                      slice_offset, slice_size, add_inputs)
+
+
+def bgmv_expand_slice(inputs: torch.Tensor,
+                      lora_b_weights: torch.Tensor,
+                      output_tensor: torch.Tensor,
+                      lora_indices_tensor: torch.Tensor,
+                      slice_offset: int,
+                      slice_size: int,
+                      add_inputs: bool = True):
+    selected_loras = lora_b_weights[lora_indices_tensor].to(
+        dtype=output_tensor.dtype)
+    inputs = inputs.to(dtype=output_tensor.dtype)
+    if len(selected_loras.shape) == 4:
+        selected_loras = selected_loras.squeeze(dim=1)
+    outputs = torch.einsum("bi, boi -> bo", inputs, selected_loras)
+
+    if add_inputs:
+        output_tensor[:, slice_offset:slice_offset + slice_size] += outputs[:]
+    else:
+        output_tensor[:, slice_offset:slice_offset + slice_size] = outputs[:]

vllm/lora/ops/triton_ops/__init__.py

@@ -0,0 +1,11 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from vllm.lora.ops.triton_ops.lora_expand import lora_expand
+from vllm.lora.ops.triton_ops.lora_kernel_metadata import LoRAKernelMeta
+from vllm.lora.ops.triton_ops.lora_shrink import lora_shrink
+
+__all__ = [
+    "lora_expand",
+    "lora_shrink",
+    "LoRAKernelMeta",
+]

vllm/lora/ops/triton_ops/kernel_utils.py

@@ -0,0 +1,243 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Utilities for Punica kernel construction.
+"""
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def mm_k(a_ptr, b_ptr, ak_stride, bk_stride, offset_k, K: tl.constexpr,
+         BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr, BLOCK_K: tl.constexpr,
+         EVEN_K: tl.constexpr, SPLIT_K: tl.constexpr, CAST_TYPE: tl.constexpr,
+         b_dtype: tl.constexpr):
+    """
+    Given a_ptr and b_ptr, that identify the rows of A (m x k) and columns of
+    B (k x n), iterate, through the K dimension to compute the partial/complete
+    matrix block product.
+    If SPLIT_K == 1, the output m x n product is complete.
+    If SPLIT_K > 1, the thread block computes partial outputs. The partial
+    outputs are then atomically summed in the caller code. 
+    Args:
+        a_ptr: Array of pointers, identifying rows of A 
+        b_ptr: Array of pointers, identifying columns of B
+        ak_stride: K dimension stride of the A matrix
+        bk_stride: K dimension stride of the B matrix
+        K: Length of the K dimension
+        BLOCK_M: M dimension of the output block m x n
+        BLOCK_N: N dimension of the output block m x n
+        BLOCK_K: K dimension atom
+        EVEN_K: True if the blocks of A and B can be loaded without any
+          masking.
+        SPLIT_K: Parameter signifying parallelism in the K dimension. 
+        CAST_TYPE: if True, cast the values from the A matrix to the B
+          matrix dtype.
+        b_dtype: datatype of the B matrix
+    """
+    accumulator = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)
+    for k in range(tl.cdiv(K, BLOCK_K * SPLIT_K)):
+        if EVEN_K:
+            tiled_a = tl.load(a_ptr)
+            tiled_b = tl.load(b_ptr)
+        else:
+            tiled_a = tl.load(a_ptr,
+                              mask=offset_k[None, :]
+                              < K - k * (BLOCK_K * SPLIT_K),
+                              other=0)
+            tiled_b = tl.load(b_ptr,
+                              mask=offset_k[:, None]
+                              < K - k * (BLOCK_K * SPLIT_K),
+                              other=0)
+        if CAST_TYPE:
+            tiled_a = tiled_a.to(b_dtype)
+        accumulator += tl.dot(
+            tiled_a,
+            tiled_b,
+        )
+        a_ptr += BLOCK_K * SPLIT_K * ak_stride
+        b_ptr += BLOCK_K * SPLIT_K * bk_stride
+    return accumulator
+
+
+@triton.jit
+def do_expand_kernel(
+    pid_n,
+    lora_index,
+    slice_id,
+    input_ptr,
+    lora_ptr,
+    out_ptr,
+    N,
+    K,
+    M_LEN,
+    ram,  # array identifying the rows of Input ptr to operate on
+    slice_start_loc,
+    # input ptr strides
+    input_d0_stride,
+    input_d1_stride,
+    input_d2_stride,
+    # lora ptr strides
+    ls_d0_ptr,
+    ls_d1_ptr,
+    ls_d2_ptr,
+    # out ptr strides
+    output_d0_stride,
+    output_d1_stride,
+    # constants
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    SAME_STRIDE: tl.constexpr,
+    SLICE_NUM: tl.constexpr,
+    EVEN_K: tl.constexpr,
+    CAST_TYPE: tl.constexpr,
+    ADD_INPUTS: tl.constexpr,
+):
+    """
+    Given an array of integers that identifies the rows of A, ram,
+    a lora index that identifies which LoRA to use from lora_ptr, lora_index,
+    a slice_id that identifies the input/output slice,
+    compute the matrix product and store in the appropriate output location.
+    Given that this is an expand kernel, we don't perform any split-K reduction
+    as the K dimension is assumed to be small.
+    """
+
+    # ls_d*_ptr can be either an integer or a pointer
+    if SAME_STRIDE:
+        # integer
+        cur_lora_d0_stride = ls_d0_ptr
+        cur_lora_d1_stride = ls_d1_ptr
+        cur_lora_d2_stride = ls_d2_ptr
+    else:
+        # pointer
+        cur_lora_d0_stride = tl.load(ls_d0_ptr + slice_id)
+        cur_lora_d1_stride = tl.load(ls_d1_ptr + slice_id)
+        cur_lora_d2_stride = tl.load(ls_d2_ptr + slice_id)
+
+    # Identify the input_ptr and lora_ptr from slice_id.
+    if SLICE_NUM == 1:
+        cur_input_ptr = input_ptr
+        cur_lora_ptr = lora_ptr
+    else:
+        cur_input_ptr = input_ptr + slice_id * input_d0_stride
+        cur_lora_ptr = tl.load(lora_ptr + slice_id).to(
+            tl.pointer_type(out_ptr.dtype.element_ty))
+
+    # Identify the column indices of B to process.
+    offset_n = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N
+    rbn = tl.max_contiguous(tl.multiple_of(offset_n % N, BLOCK_N), BLOCK_N)
+
+    # Identify A and B block pointers
+    offset_k = tl.arange(0, BLOCK_K)
+    a_ptr = (cur_input_ptr + ram[:, None] * input_d1_stride +
+             offset_k[None, :] * input_d2_stride)
+    b_ptr = (cur_lora_ptr + cur_lora_d0_stride * lora_index +
+             offset_k[:, None] * cur_lora_d2_stride +
+             rbn[None, :] * cur_lora_d1_stride)
+
+    # Compute the block matrix product.
+    SPLIT_K = 1
+    accumulator = mm_k(a_ptr, b_ptr, input_d2_stride, cur_lora_d2_stride,
+                       offset_k, K, BLOCK_M, BLOCK_N, BLOCK_K, EVEN_K, SPLIT_K,
+                       CAST_TYPE, cur_lora_ptr.dtype.element_ty)
+
+    tiled_c = accumulator.to(cur_lora_ptr.dtype.element_ty)
+    if SLICE_NUM == 1:
+        cur_slice_start = slice_start_loc
+    else:
+        cur_slice_start = tl.load(slice_start_loc + slice_id)
+
+    # Identify the C output pointers to store the results of the accumulator.
+    offset_cn = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N + cur_slice_start
+    offset_cm = tl.arange(0, BLOCK_M)
+    c_ptr = (out_ptr + ram[:, None] * output_d0_stride +
+             offset_cn[None, :] * output_d1_stride)
+    c_mask = (offset_cm[:, None] < M_LEN) & (offset_cn[None, :]
+                                             < (cur_slice_start + N))
+
+    if ADD_INPUTS:
+        tiled_out = tl.load(c_ptr, mask=c_mask)
+        tiled_c += tiled_out
+    tl.store(c_ptr, tiled_c, mask=c_mask)
+
+
+@triton.jit
+def do_shrink_kernel(
+    pid_n,
+    pid_sk,
+    slice_id,
+    lora_index,
+    input_ptr,
+    lora_ptr,
+    out_ptr,
+    N,
+    K,
+    M_LEN,
+    ram,
+    # input strides
+    input_d0_stride,
+    input_d1_stride,
+    # lora strides
+    lora_d0_stride,
+    lora_d1_stride,
+    lora_d2_stride,
+    # output strides
+    output_d0_stride,
+    output_d1_stride,
+    output_d2_stride,
+    scaling,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_K: tl.constexpr,
+    EVEN_K: tl.constexpr,
+    SPLIT_K: tl.constexpr,
+    SLICE_NUM: tl.constexpr,
+):
+    """
+    Given an array of integers that identifies the rows of A, ram,
+    a lora index that identifies which LoRA to use from lora_ptr, lora_index,
+    a slice_id that identifies the input/output slice, compute the
+    matrix product and store in the appropriate output location.
+    """
+
+    # Identify the lora_ptr from slice_id.
+    if SLICE_NUM == 1:
+        # current lora ptr
+        cur_lora_ptr = lora_ptr
+    else:
+        # current lora ptr
+        cur_lora_ptr = tl.load(lora_ptr + slice_id).to(
+            tl.pointer_type(input_ptr.dtype.element_ty))
+
+    # Identify the column indices of B to process.
+    offset_n = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N
+    rbn = tl.max_contiguous(tl.multiple_of(offset_n % N, BLOCK_N), BLOCK_N)
+
+    # Identify A and B block pointers
+    offset_k = pid_sk * BLOCK_K + tl.arange(0, BLOCK_K)
+    a_ptr = (input_ptr + ram[:, None] * input_d0_stride +
+             offset_k[None, :] * input_d1_stride)
+    b_ptr = (cur_lora_ptr + lora_d0_stride * lora_index +
+             rbn[None, :] * lora_d1_stride +
+             offset_k[:, None] * lora_d2_stride)
+
+    # Compute partial/complete block matrix product.
+    accumulator = mm_k(a_ptr, b_ptr, input_d1_stride, lora_d2_stride, offset_k,
+                       K, BLOCK_M, BLOCK_N, BLOCK_K, EVEN_K, SPLIT_K, False,
+                       cur_lora_ptr.dtype.element_ty)
+
+    # Identify the C output pointers to store the results of the accumulator.
+    offset_cn = tl.arange(0, BLOCK_N) + pid_n * BLOCK_N
+    offset_cm = tl.arange(0, BLOCK_M)
+    cur_out_ptr = (out_ptr if SLICE_NUM == 1 else out_ptr +
+                   slice_id * output_d0_stride)
+    c_ptr = cur_out_ptr + ram[:, None] * output_d1_stride + offset_cn[
+        None, :] * output_d2_stride
+    c_mask = (offset_cm[:, None] < M_LEN) & (offset_cn[None, :] < N)
+
+    accumulator *= scaling
+    # handles write-back with reduction-splitting
+    if SPLIT_K == 1:
+        tl.store(c_ptr, accumulator, mask=c_mask)
+    else:
+        tl.atomic_add(c_ptr, accumulator, mask=c_mask)

vllm/lora/ops/triton_ops/lora_expand.py

@@ -0,0 +1,293 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Based on:
+Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023).
+Punica: Multi-Tenant LoRA Serving.
+https://arxiv.org/abs/2310.18547
+"""
+
+from typing import List
+
+import torch
+import triton
+import triton.language as tl
+
+from vllm.lora.ops.triton_ops.kernel_utils import do_expand_kernel
+from vllm.lora.ops.triton_ops.utils import _get_lora_b_ptr
+from vllm.utils import direct_register_custom_op
+
+
+@triton.jit
+def _lora_expand_kernel(
+        input_ptr,
+        lora_ptr,
+        out_ptr,
+        M,
+        N,
+        K,
+        token_indices_sorted_by_lora_ids,
+        num_tokens_per_lora,
+        lora_token_start_loc,
+        lora_ids,
+        slice_start_loc,
+        input_d0_stride,
+        input_d1_stride,
+        input_d2_stride,  # 1
+        ls_d0_ptr,
+        ls_d1_ptr,
+        ls_d2_ptr,  # 1
+        output_d0_stride,
+        output_d1_stride,  # 1
+        output_hs_ptr,
+        BLOCK_M: tl.constexpr,
+        BLOCK_N: tl.constexpr,
+        BLOCK_K: tl.constexpr,
+        EVEN_K: tl.constexpr,
+        ADD_INPUTS: tl.constexpr,
+        CAST_TYPE: tl.constexpr,
+        SLICE_NUM: tl.constexpr,
+        SAME_STRIDE: tl.constexpr):
+
+    cta_n_num = tl.cdiv(N, BLOCK_N)
+    cta_m_num = tl.cdiv(M, BLOCK_M)
+
+    pid_mn = tl.program_id(axis=0)
+    pid_m = pid_mn % cta_m_num
+    pid_n = (pid_mn // cta_m_num) % cta_n_num
+
+    slice_id = tl.program_id(axis=1)
+    lora_idx = tl.program_id(axis=2)
+
+    lora_id = tl.load(lora_ids + lora_idx)
+    if lora_id == -1:
+        # Early exit for the no-lora case.
+        return
+
+    lora_m_size = tl.load(num_tokens_per_lora + lora_idx)
+
+    cta_m_offset = pid_m * BLOCK_M
+    if cta_m_offset >= lora_m_size:
+        # Early exit CTA.
+        return
+
+    # When the output dimensions of each slice are the same,cur_n=N, otherwise
+    # cur_n=tl.load(output_hs_ptr + slice_id), this situation exists in GQA's
+    # qkv linear.
+    curr_N = N if SAME_STRIDE else tl.load(output_hs_ptr + slice_id)
+    if pid_n * BLOCK_N >= curr_N:
+        # Early exit CTA.
+        return
+
+    # num rows this CTA should process.
+    cta_m_len = min(BLOCK_M, lora_m_size - cta_m_offset)
+
+    # Identify all rows that this CTA should process.
+    lora_m_indices_start = tl.load(lora_token_start_loc + lora_idx)
+    cta_lora_seq_indices = (token_indices_sorted_by_lora_ids +
+                            lora_m_indices_start + cta_m_offset)
+
+    # Load all relevant row indices.
+    offset_m = tl.arange(0, BLOCK_M) % cta_m_len
+    ram = tl.load(cta_lora_seq_indices + offset_m)
+
+    do_expand_kernel(
+        pid_n,
+        lora_id,
+        slice_id,
+        input_ptr,
+        lora_ptr,
+        out_ptr,
+        curr_N,
+        K,
+        cta_m_len,
+        ram,  # array identifying the rows of Input ptr to operate on
+        slice_start_loc,
+        # input ptr strides
+        input_d0_stride,
+        input_d1_stride,
+        input_d2_stride,
+        # lora ptr strides
+        ls_d0_ptr,
+        ls_d1_ptr,
+        ls_d2_ptr,
+        # out ptr strides
+        output_d0_stride,
+        output_d1_stride,
+        # constants
+        BLOCK_M,
+        BLOCK_N,
+        BLOCK_K,
+        SAME_STRIDE,
+        SLICE_NUM,
+        EVEN_K,
+        CAST_TYPE,
+        ADD_INPUTS)
+
+
+@torch.inference_mode()
+def _lora_expand(
+    inputs: torch.Tensor,  # shape [num_slices, num_tokens, lora_rank]
+    lora_b_weights: List[
+        torch.Tensor],  # shape [num_lora, hidden_size, lora_rank]
+    output_tensor: torch.
+    Tensor,  # shape [num_tokens, hidden_size * num_slices]
+    token_lora_mapping: torch.Tensor,  # shape [num_tokens]
+    token_indices_sorted_by_lora_ids: torch.Tensor,  # shape [num_tokens]
+    num_tokens_per_lora: torch.Tensor,  # shape [max-loras + 1]
+    lora_token_start_loc: torch.Tensor,  # shape [max-loras + 2]
+    lora_ids: torch.Tensor,  # shape [max-loras + 1]
+    no_lora_flag_cpu: torch.Tensor,  # shape [1] 
+    offset_start: int = 0,
+    add_inputs: bool = False,
+) -> None:
+    """
+    Args:
+        inputs (torch.Tensor): input tensor
+        lora_b_weights (List[torch.Tensor]): lora'b weight
+        output_tensor (torch.Tensor): output tensor
+        token_lora_mapping (torch.Tensor): A tensor mapping each input token
+            to the lora-id related to that token. A value of -1 indicates that
+            LoRA doesn't apply to that token.
+        token_indices_sorted_by_lora_ids (torch.Tensor): Row/Token indices from
+            the A matrix grouped by LoRA IDs.
+        num_tokens_per_lora (torch.Tensor): num_tokens_per_lora[i] is the number
+            of tokens that are to be processed by LoRA ID lora_ids[i] 
+        lora_token_start_loc (torch.Tensor): A cumulative sum of
+            num_tokens_per_lora. lora_token_start_loc[0] is always 0 so that
+            lora_token_start_loc[i], along with num_tokens_per_lora[i]
+            identifies the the region in token_indices_sorted_by_lora_ids that
+            LoRA lora_ids[i] should process.
+        lora_ids (torch.Tensor): LoRA ids to process.
+        no_lora_flag_cpu (torch.Tensor): A CPU tensor of size 1, that indicates
+            if there are any requests that require LoRA.
+        offset_start (int, optional): Offset start for output_tensor. 
+            Defaults to 0.
+        add_inputs (bool, optional): Whether to add the input tensor to the 
+            output tensor. Defaults to False.
+    """
+
+    assert no_lora_flag_cpu.numel() == 1
+    if no_lora_flag_cpu.item():
+        # None of the inputs require LoRA.
+        return
+
+    assert inputs.dtype in [torch.float16, torch.bfloat16, torch.float32]
+    for weight in lora_b_weights:
+        assert weight.dtype in [torch.float16, torch.bfloat16]
+
+    assert inputs.size(0) == len(lora_b_weights)
+    assert output_tensor.is_contiguous()
+
+    # metadata sanity check.
+    M = inputs.size(1)
+    assert token_lora_mapping.size(0) == M
+    assert token_lora_mapping.size(0) == token_indices_sorted_by_lora_ids.size(
+        0)
+    assert lora_ids.size(0) == num_tokens_per_lora.size(0)
+    assert lora_token_start_loc.size(0) == lora_ids.size(0) + 1
+
+    (slice_start_tensor, lora_ptr_tensor, lora_strides_d0_tensor,
+     lora_strides_d1_tensor, lora_strides_d2_tensor, hidden_sizes_tensor,
+     same_stride, MAX_N) = _get_lora_b_ptr(lora_b_weights, offset_start,
+                                           inputs.device)
+
+    K = lora_b_weights[0].shape[-1]  # K= rank
+    ADD_INPUTS = add_inputs
+    MAX_LORAS = lora_ids.size(0)
+    CAST_TYPE = False
+    NUM_SLICES = len(lora_b_weights)
+
+    # Triton kernel configs.
+    BLOCK_M = 64
+    BLOCK_N = 128
+    BLOCK_K = 16
+    NUM_WARPS = 4
+    NUM_CTAS = 1
+    NUM_STAGES = 2
+    MAX_NREG = None
+
+    EVEN_K = K % BLOCK_K == 0  # type: ignore
+
+    if inputs.dtype == torch.float32 and lora_b_weights[0].dtype in [
+            torch.float16,
+            torch.bfloat16,
+    ]:
+        CAST_TYPE = True
+
+    # TODO (varun): This grid formulation maximizes parallelization at the
+    # cost of wasteful thread block launch when only a few input tokens require
+    # LoRA. This might not be the best in all cases.
+    grid = (
+        triton.cdiv(M, BLOCK_M) * triton.cdiv(MAX_N, BLOCK_N),
+        NUM_SLICES,
+        # Each LoRA receives its own set of thread blocks for output
+        # computation. If some LoRA doesn't have any tokens to process, its
+        # thread blocks simply exit.
+        MAX_LORAS,
+    )
+
+    _lora_expand_kernel[grid](
+        inputs,
+        lora_ptr_tensor,
+        output_tensor,
+        M,
+        MAX_N,
+        K,
+        token_indices_sorted_by_lora_ids,
+        num_tokens_per_lora,
+        lora_token_start_loc,
+        lora_ids,
+        slice_start_tensor,
+        inputs.stride(0),
+        inputs.stride(1),
+        inputs.stride(2),
+        lora_strides_d0_tensor,
+        lora_strides_d1_tensor,
+        lora_strides_d2_tensor,
+        output_tensor.stride(0),
+        output_tensor.stride(1),
+        hidden_sizes_tensor,
+        BLOCK_M,
+        BLOCK_N,
+        BLOCK_K,
+        EVEN_K,
+        ADD_INPUTS,
+        CAST_TYPE,
+        NUM_SLICES,
+        same_stride,
+        num_warps=NUM_WARPS,
+        num_ctas=NUM_CTAS,
+        num_stages=NUM_STAGES,
+        maxnreg=MAX_NREG,
+    )
+
+    return
+
+
+def _lora_expand_fake(
+    inputs: torch.Tensor,
+    lora_b_weights: List[torch.Tensor],
+    output_tensor: torch.Tensor,
+    token_lora_mapping: torch.Tensor,
+    token_indices_sorted_by_lora_ids: torch.Tensor,
+    num_tokens_per_lora: torch.Tensor,
+    lora_token_start_loc: torch.Tensor,
+    lora_ids: torch.Tensor,
+    no_lora_flag_cpu: torch.Tensor,
+    offset_start: int = 0,
+    add_inputs: bool = False,
+) -> None:
+    return
+
+
+try:
+    direct_register_custom_op(
+        op_name="lora_expand",
+        op_func=_lora_expand,
+        mutates_args=["output_tensor"],
+        fake_impl=_lora_expand_fake,
+    )
+    lora_expand = torch.ops.vllm.lora_expand
+
+except AttributeError:
+    lora_expand = _lora_expand

vllm/lora/ops/triton_ops/lora_kernel_metadata.py

@@ -0,0 +1,147 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+LoRA kernels metadata preparation utilities.
+"""
+
+from dataclasses import dataclass
+from typing import Tuple, Union
+
+import torch
+
+
+@dataclass
+class LoRAKernelMeta:
+    token_lora_mapping: torch.Tensor
+    token_indices_sorted_by_lora_ids: torch.Tensor
+    active_lora_ids: torch.Tensor
+    num_tokens_per_lora: torch.Tensor
+    lora_token_start_loc: torch.Tensor
+
+    # The V1 architecture uses the traced torch.compile graphs to execute
+    # a forward pass. Things to note about this process,
+    # 1. The tracing infers all python scalar datatype objects into a constant
+    # value.
+    # 2. The tracing cannot handle dynamic control flow. (dynamic control flow
+    # is an experimental feature in pytorch)
+    # 3. The internals of torch.ops functions are not traced.
+    # We disguise the "no_lora" flag as a cpu tensor and leverage point number 3
+    # to early exit from inside the lora_expand / lora_shrink torch operation.
+    no_lora_flag_cpu: torch.Tensor
+
+    @staticmethod
+    def make(max_loras: int, max_num_tokens: int,
+             device: Union[torch.device, str]) -> "LoRAKernelMeta":
+
+        token_lora_mapping = torch.empty(max_num_tokens,
+                                         dtype=torch.int32,
+                                         device=device)
+
+        token_indices_sorted_by_lora_ids = torch.empty(max_num_tokens,
+                                                       dtype=torch.int32,
+                                                       device=device)
+
+        # +1 because "no-lora" is also a possibility
+        # example: let max_loras be 3, active_lora_ids of [-1, 0, 2, 1]
+        # is a possibility.
+        active_lora_ids = torch.empty(max_loras + 1,
+                                      dtype=torch.int32,
+                                      device=device)
+
+        # using running example, [3, 10, 5, 2] is a possibility.
+        num_tokens_per_lora = torch.zeros(max_loras + 1,
+                                          dtype=torch.int32,
+                                          device=device)
+
+        # +2 for this because, the first index is always 0.
+        # using running example, lora_token_start_loc
+        # is [0, 3, 13, 18, 20].
+        lora_token_start_loc = torch.zeros(max_loras + 2,
+                                           dtype=torch.int32,
+                                           device=device)
+
+        no_lora_flag_cpu = torch.tensor([False],
+                                        dtype=torch.bool,
+                                        device='cpu')
+
+        return LoRAKernelMeta(
+            token_lora_mapping=token_lora_mapping,
+            token_indices_sorted_by_lora_ids=token_indices_sorted_by_lora_ids,
+            active_lora_ids=active_lora_ids,
+            num_tokens_per_lora=num_tokens_per_lora,
+            lora_token_start_loc=lora_token_start_loc,
+            no_lora_flag_cpu=no_lora_flag_cpu)
+
+    def _reset(self):
+        self.active_lora_ids.fill_(-1)
+        self.num_tokens_per_lora.fill_(0)
+        self.lora_token_start_loc.fill_(0)
+        self.no_lora_flag_cpu.fill_(False)
+
+    def prepare_tensors(self, token_lora_mapping: torch.Tensor) -> None:
+        """
+        Prepare kernel metadata tensors for the current forward pass.
+
+        Args:
+            token_lora_tensor (torch.Tensor): Tensor containing lora indices
+            for each input token.
+        """
+
+        self._reset()
+
+        # Check and record no-lora case.
+        no_lora = torch.all(token_lora_mapping == -1)
+        self.no_lora_flag_cpu[0] = no_lora
+
+        if no_lora:
+            # Early exit. LoRA kernels will not be run.
+            return
+
+        num_tokens = token_lora_mapping.size(0)
+
+        # copy token lora mapping
+        self.token_lora_mapping[:num_tokens].copy_(token_lora_mapping,
+                                                   non_blocking=True)
+
+        # token_indices_sorted_by_lora_ids
+        _, token_indices_sorted_by_lora_ids = torch.sort(token_lora_mapping,
+                                                         stable=True)
+        # start gpu transfer
+        self.token_indices_sorted_by_lora_ids[:num_tokens].copy_(
+            token_indices_sorted_by_lora_ids, non_blocking=True)
+
+        # active_lora_ids, num_tokens_per_lora
+        lora_ids, num_tokens_per_lora = torch.unique(token_lora_mapping,
+                                                     sorted=False,
+                                                     return_counts=True)
+        self.active_lora_ids[:lora_ids.size(0)].copy_(lora_ids,
+                                                      non_blocking=True)
+        self.num_tokens_per_lora[:num_tokens_per_lora.size(0)].copy_(
+            num_tokens_per_lora, non_blocking=True)
+
+        # lora_token_start_loc
+        lora_token_start_loc = torch.cumsum(num_tokens_per_lora, dim=0)
+        self.lora_token_start_loc[1:1 + lora_token_start_loc.size(0)].copy_(
+            lora_token_start_loc, non_blocking=True)
+
+    def meta_args(
+        self, token_nums: int
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor,
+               torch.Tensor, torch.Tensor]:
+        """
+        This function returns the kernel metadata required for the current
+        forward pass execution of the kernel. The function returns all the
+        metadata required by the kernel, in order, as a tuple, so it can be
+        unpacked directly during the lora_shrink/lora_expand function call.
+
+        Args:
+            token_nums (int): Number of input tokens in the current forward
+            pass. 
+        """
+        return (
+            self.token_lora_mapping[:token_nums],
+            self.token_indices_sorted_by_lora_ids[:token_nums],
+            self.num_tokens_per_lora,
+            self.lora_token_start_loc,
+            self.active_lora_ids,
+            self.no_lora_flag_cpu,
+        )

vllm/lora/ops/triton_ops/lora_shrink.py

@@ -0,0 +1,247 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+Based on:
+Chen, L., Ye, Z., Wu, Y., Zhuo, D., Ceze, L., & Krishnamurthy, A. (2023). 
+Punica: Multi-Tenant LoRA Serving. 
+https://arxiv.org/abs/2310.18547
+"""
+
+from typing import List
+
+import torch
+import triton
+import triton.language as tl
+
+from vllm.lora.ops.triton_ops.kernel_utils import do_shrink_kernel
+from vllm.lora.ops.triton_ops.utils import _get_lora_a_ptr
+from vllm.utils import direct_register_custom_op
+
+
+@triton.jit
+def _lora_shrink_kernel(input_ptr, lora_ptr, out_ptr, M, N, K,
+                        token_indices_sorted_by_lora_ids, num_tokens_per_lora,
+                        lora_token_start_loc, lora_ids, scaling,
+                        input_d0_stride, input_d1_stride, lora_d0_stride,
+                        lora_d1_stride, lora_d2_stride, output_d0_stride,
+                        output_d1_stride, output_d2_stride,
+                        BLOCK_M: tl.constexpr, BLOCK_N: tl.constexpr,
+                        BLOCK_K: tl.constexpr, EVEN_K: tl.constexpr,
+                        SPLIT_K: tl.constexpr, SLICE_NUM: tl.constexpr):
+
+    cta_n_num = tl.cdiv(N, BLOCK_N)
+    cta_m_num = tl.cdiv(M, BLOCK_M)
+
+    pid_sk_m_n = tl.program_id(axis=0)
+    pid_sk = pid_sk_m_n % SPLIT_K
+    pid_m = (pid_sk_m_n // SPLIT_K) % cta_m_num
+    pid_n = pid_sk_m_n // (SPLIT_K * cta_m_num) % cta_n_num
+
+    slice_id = tl.program_id(axis=1)
+    lora_idx = tl.program_id(axis=2)
+
+    lora_id = tl.load(lora_ids + lora_idx)
+    if lora_id == -1:
+        # Early exit for the no-lora case.
+        return
+
+    lora_m_size = tl.load(num_tokens_per_lora + lora_idx)
+
+    cta_m_offset = pid_m * BLOCK_M
+    if cta_m_offset >= lora_m_size:
+        # Early exit CTA.
+        return
+
+    # num rows this CTA should process.
+    cta_m_len = min(BLOCK_M, lora_m_size - cta_m_offset)
+
+    # Identify all rows that this CTA should process.
+    lora_m_indices_start = tl.load(lora_token_start_loc + lora_idx)
+    cta_lora_seq_indices = (token_indices_sorted_by_lora_ids +
+                            lora_m_indices_start + cta_m_offset)
+
+    # Load all relevant row indices.
+    offset_m = tl.arange(0, BLOCK_M) % cta_m_len
+    ram = tl.load(cta_lora_seq_indices + offset_m)
+
+    do_shrink_kernel(
+        pid_n,
+        pid_sk,
+        slice_id,
+        lora_id,
+        input_ptr,
+        lora_ptr,
+        out_ptr,
+        N,
+        K,
+        cta_m_len,
+        ram,  # array identifying the rows of Input ptr to operate on
+        # input strides
+        input_d0_stride,
+        input_d1_stride,
+        # lora strides
+        lora_d0_stride,
+        lora_d1_stride,
+        lora_d2_stride,
+        # output strides
+        output_d0_stride,
+        output_d1_stride,
+        output_d2_stride,
+        scaling,
+        BLOCK_M,
+        BLOCK_N,
+        BLOCK_K,
+        EVEN_K,
+        SPLIT_K,
+        SLICE_NUM)
+
+
+@torch.inference_mode()
+def _lora_shrink(
+    inputs: torch.Tensor,  #  shape [num_tokens, hidden_size]
+    lora_a_weights: List[
+        torch.Tensor],  # shape [num_loras, lora_rank, hidden_size]
+    output_tensor: torch.Tensor,  # shape [num_slices, num_tokens, lora_rank]
+    token_lora_mapping: torch.Tensor,  # shape [num_tokens]
+    token_indices_sorted_by_lora_ids: torch.Tensor,  # shape [num_tokens] 
+    num_tokens_per_lora: torch.Tensor,  # shape [max-loras + 1]
+    lora_token_start_loc: torch.Tensor,  # shape [max-loras + 2]
+    lora_ids: torch.Tensor,  # shape [max-loras + 1]
+    no_lora_flag_cpu: torch.Tensor,  # shape [1]
+    scaling: float,
+) -> None:
+    """
+    Args:
+        inputs (torch.Tensor): Input tensor
+        lora_a_weights (List[torch.Tensor]): LoRA weights
+        output_tensor (torch.Tensor): output tensor
+        token_lora_mapping (torch.Tensor): A tensor mapping each input token
+            to the lora-id related to that token. A value of -1 indicates that
+            LoRA doesn't apply to that token.
+        token_indices_sorted_by_lora_ids (torch.Tensor): Row/Token indices from
+            the A matrix grouped by LoRA IDs.
+        num_tokens_per_lora (torch.Tensor): num_tokens_per_lora[i] is the number
+            of tokens that are to be processed by LoRA ID lora_ids[i] 
+        lora_token_start_loc (torch.Tensor): A cumulative sum of
+            num_tokens_per_lora. lora_token_start_loc[0] is always 0 so that
+            lora_token_start_loc[i], along with num_tokens_per_lora[i]
+            identifies the region in token_indices_sorted_by_lora_ids that
+            LoRA lora_ids[i] should process.
+        lora_ids (torch.Tensor): LoRA ids to process.
+        no_lora_flag_cpu (torch.Tensor): A CPU tensor of size 1, that indicates
+            if there are any requests that require LoRA.
+        scaling (float): Scaling factor.
+    """
+
+    assert no_lora_flag_cpu.numel() == 1
+    if no_lora_flag_cpu.item():
+        # None of the inputs require LoRA.
+        return
+
+    assert inputs.dtype == lora_a_weights[0].dtype
+    assert inputs.dtype in [torch.float16, torch.bfloat16]
+    for weight in lora_a_weights:
+        assert weight.dtype in [torch.float16, torch.bfloat16]
+
+    assert inputs.size(1) == lora_a_weights[0].size(-1)
+    assert inputs.is_contiguous()
+    assert output_tensor.is_contiguous()
+
+    # metadata sanity check
+    M = inputs.size(0)
+    assert token_lora_mapping.size(0) == M
+    assert token_lora_mapping.size(0) == token_indices_sorted_by_lora_ids.size(
+        0)
+    assert lora_ids.size(0) == num_tokens_per_lora.size(0)
+    assert lora_token_start_loc.size(0) == lora_ids.size(0) + 1
+
+    (lora_ptr_tensor, lora_strides_d0, lora_strides_d1,
+     lora_strides_d2) = _get_lora_a_ptr(lora_a_weights, inputs.device)
+    N, K = lora_a_weights[0].shape[-2:]  # K=hidden_size,N=rank
+    NUM_SLICES = len(lora_a_weights)
+    MAX_LORAS = lora_ids.size(0)
+
+    # Triton kernel configs
+    BLOCK_M = 32
+    BLOCK_N = 16
+    BLOCK_K = 256 if M < 128 else 32
+    SPLIT_K = 64 if M < 128 else 8
+    NUM_WARPS = 4
+    NUM_CTAS = 1
+    NUM_STAGES = 2
+    MAX_NREG = None
+
+    EVEN_K = K % (BLOCK_K * SPLIT_K) == 0  # type: ignore
+
+    # TODO (varun): This grid formulation maximizes parallelization at the
+    # cost of wasteful thread block launch when only few of the input tokens
+    # require LoRA. This might not be the best in all cases.
+    grid = (
+        SPLIT_K * triton.cdiv(M, BLOCK_M) * triton.cdiv(N, BLOCK_N),
+        NUM_SLICES,
+        # Each LoRA receives its own set of thread blocks for output
+        # computation. If some LoRA doesn't have any tokens to process, its
+        # thread blocks exit early.
+        MAX_LORAS,
+    )
+
+    _lora_shrink_kernel[grid](
+        inputs,
+        lora_ptr_tensor,
+        output_tensor,
+        M,
+        N,
+        K,
+        token_indices_sorted_by_lora_ids,
+        num_tokens_per_lora,
+        lora_token_start_loc,
+        lora_ids,
+        scaling,
+        inputs.stride(0),
+        inputs.stride(1),
+        lora_strides_d0,
+        lora_strides_d1,
+        lora_strides_d2,
+        output_tensor.stride(0),
+        output_tensor.stride(1),
+        output_tensor.stride(2),
+        BLOCK_M,
+        BLOCK_N,
+        BLOCK_K,
+        EVEN_K,
+        SPLIT_K,
+        NUM_SLICES,
+        num_warps=NUM_WARPS,
+        num_ctas=NUM_CTAS,
+        num_stages=NUM_STAGES,
+        maxnreg=MAX_NREG,
+    )
+
+    return
+
+
+def _lora_shrink_fake(
+    inputs: torch.Tensor,
+    lora_a_weights: List[torch.Tensor],
+    output_tensor: torch.Tensor,
+    token_lora_mapping: torch.Tensor,
+    token_indices_sorted_by_lora_ids: torch.Tensor,
+    num_tokens_per_lora: torch.Tensor,
+    lora_token_start_loc: torch.Tensor,
+    lora_ids: torch.Tensor,
+    no_lora_flag_cpu: torch.Tensor,
+    scaling: float,
+) -> None:
+    return
+
+
+try:
+    direct_register_custom_op(
+        op_name="lora_shrink",
+        op_func=_lora_shrink,
+        mutates_args=["output_tensor"],
+        fake_impl=_lora_shrink_fake,
+    )
+    lora_shrink = torch.ops.vllm.lora_shrink
+
+except AttributeError:
+    lora_shrink = _lora_shrink

vllm/lora/ops/triton_ops/utils.py

@@ -0,0 +1,121 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Dict, List, Tuple
+
+import torch
+
+_LORA_A_PTR_DICT: Dict[Tuple[int, ...], Tuple[torch.tensor, ...]] = {}
+_LORA_B_PTR_DICT: Dict[Tuple[int, ...], Tuple[torch.tensor, ...]] = {}
+
+
+def _get_lora_a_ptr(lora_a_weights: List[torch.Tensor], device: torch.device):
+    """
+    `_LORA_A_PTR_DICT` collects the required information during `profile_run`, 
+    After this, it remains constant and subsequent usage is through LUT.
+    Refer to: 
+    https://github.com/triton-lang/triton/blob/release/3.1.x/python/tutorials/08-grouped-gemm.py
+    """
+    key = tuple(lora_weight.data_ptr() for lora_weight in lora_a_weights)
+
+    if values := _LORA_A_PTR_DICT.get(key):
+        return values
+
+    lora_strides_d0 = []
+    lora_strides_d1 = []
+    lora_strides_d2 = []
+    tensor_ptrs = []
+    for lora_a_weight in lora_a_weights:
+        if lora_a_weight.ndim == 4:  # shape:(lora_num,1,size,rank)
+            assert lora_a_weight.size(1) == 1
+            lora_a_weight = lora_a_weight.squeeze(dim=1)
+        else:
+            assert lora_a_weight.ndim == 3  # shape:(lora_num,size,rank)
+        assert lora_a_weight.is_contiguous()
+        tensor_ptrs.append(lora_a_weight.data_ptr())
+        lora_strides_d0.append(lora_a_weight.stride(0))
+        lora_strides_d1.append(lora_a_weight.stride(1))
+        lora_strides_d2.append(lora_a_weight.stride(2))
+    if len(lora_a_weights) > 1:
+        lora_ptr_tensor = torch.tensor(tensor_ptrs, device=device)
+    else:
+        lora_ptr_tensor = lora_a_weights[0]
+
+    if (len(set(lora_strides_d0)) > 1 or len(set(lora_strides_d1)) > 1
+            or len(set(lora_strides_d2)) > 1):
+        raise ValueError("All LoRA weights must have the same stride.")
+
+    _LORA_A_PTR_DICT[key] = (
+        lora_ptr_tensor,
+        lora_strides_d0[0],
+        lora_strides_d1[0],
+        lora_strides_d2[0],
+    )
+    return _LORA_A_PTR_DICT.get(key)
+
+
+def _get_lora_b_ptr(lora_weights: List[torch.Tensor], offset_start: int,
+                    device: torch.device):
+    """ 
+     `_LORA_B_PTR_DICT` collects the required information during `profile_run`, 
+    After this, it remains constant and subsequent usage is through LUT.
+    Refer to: 
+    https://github.com/triton-lang/triton/blob/release/3.1.x/python/tutorials/08-grouped-gemm.py
+
+    """
+
+    key = tuple(lora_weight.data_ptr() for lora_weight in lora_weights)
+    if values := _LORA_B_PTR_DICT.get(key):
+        return values
+    slice_offset_lst = []
+    tensor_ptrs = []
+    lora_strides_d0 = []
+    lora_strides_d1 = []
+    lora_strides_d2 = []
+    hidden_sizes = []
+    slice_offset = offset_start
+    for lora_b_weight in lora_weights:
+        if lora_b_weight.ndim == 4:  # shape:(lora_num,1,size,rank)
+            assert lora_b_weight.size(1) == 1
+            lora_b_weight = lora_b_weight.squeeze(dim=1)
+        else:
+            assert lora_b_weight.ndim == 3  # shape:(lora_num,size,rank)
+        assert lora_b_weight.is_contiguous()
+        tensor_ptrs.append(lora_b_weight.data_ptr())
+        lora_strides_d0.append(lora_b_weight.stride(0))
+        lora_strides_d1.append(lora_b_weight.stride(1))
+        lora_strides_d2.append(lora_b_weight.stride(2))
+        slice_offset_lst.append(slice_offset)
+        slice_offset += lora_b_weight.size(1)
+        hidden_sizes.append(lora_b_weight.size(1))
+
+    if len(lora_weights) > 1:
+        # note these are device tensors
+        lora_ptr_tensor = torch.tensor(tensor_ptrs, device=device)
+        slice_start_tensor = torch.tensor(slice_offset_lst, device=device)
+    else:
+        slice_start_tensor = slice_offset_lst[0]
+        lora_ptr_tensor = lora_b_weight[0]
+
+    # If each lora has the same stride, there's no need to use a
+    # tensor for storage.
+    if (len(set(lora_strides_d0)) == 1 and len(set(lora_strides_d1)) == 1 and
+            len(set(lora_strides_d2)) == 1) and len(set(hidden_sizes)) == 1:
+        lora_strides_d0_tensor = lora_strides_d0[0]
+        lora_strides_d1_tensor = lora_strides_d1[0]
+        lora_strides_d2_tensor = lora_strides_d2[0]
+        hidden_sizes_tensor = hidden_sizes[0]
+        same_stride = True
+
+    else:
+        lora_strides_d0_tensor = torch.tensor(lora_strides_d0, device=device)
+        lora_strides_d1_tensor = torch.tensor(lora_strides_d1, device=device)
+        lora_strides_d2_tensor = torch.tensor(lora_strides_d2, device=device)
+        hidden_sizes_tensor = torch.tensor(hidden_sizes, device=device)
+        same_stride = False
+    # MAX_N is the maximum hidden size among all the lora_b weights
+    MAX_N = max(hidden_sizes)
+    _LORA_B_PTR_DICT[key] = (slice_start_tensor, lora_ptr_tensor,
+                             lora_strides_d0_tensor, lora_strides_d1_tensor,
+                             lora_strides_d2_tensor, hidden_sizes_tensor,
+                             same_stride, MAX_N)
+    return _LORA_B_PTR_DICT.get(key)

vllm/lora/ops/utils.py

@@ -0,0 +1,46 @@
+import functools
+from typing import Dict
+
+
+@functools.lru_cache
+def _get_op_configs(op_type: str, batch: int, hidden_size: int):
+    # TODO: add optimal configurations
+    return None
+
+
+def _check_divisibility(hidden_size: int):
+    # The bgmv_expand kernel requires that the hidden_size be divisible by
+    # the number below.
+    divisibility = [2, 4, 8, 16, 32, 64]
+    divisibility.sort(reverse=True)
+    for div in divisibility:
+        if hidden_size % div == 0:
+            return div
+    # hidden_size is an odd number
+    return 1
+
+
+def _get_default_config(op_type: str, batch: int, hidden_size: int):
+    if op_type == "expand":
+        return {
+            "BLOCK_N": 256,
+            "SPLIT_N": _check_divisibility(hidden_size),
+            "num_warps": 8
+        }
+    else:
+        return {"BLOCK_K": 256, "SPLIT_K": 64, "num_warps": 8}
+
+
+def get_lora_op_configs(op_type: str, batch: int,
+                        hidden_size: int) -> Dict[str, int]:
+    """Inspired by `fused_moe_kernel`
+    The return value will be a dictionary mapping an irregular grid of batch 
+    sizes and hidden_size to configurations of the bgmv-related kernel. 
+    NOTE: It currently only supports the default configuration. We plan to 
+    generate optimal configurations for different hardware in the future using 
+    scripts similar to `benchmark_moe.py`.
+    """
+    config = _get_op_configs(op_type, batch, hidden_size)
+    if not config:
+        config = _get_default_config(op_type, batch, hidden_size)
+    return config