vllm-cpu 0.11.0.post2__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/model_executor/layers/fla/ops/solve_tril.py

@@ -0,0 +1,365 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
+#
+# This file contains code copied from the flash-linear-attention project.
+# The original source code was licensed under the MIT license and included
+# the following copyright notice:
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+# ruff: noqa: E501
+from typing import Optional
+
+import torch
+
+from vllm.triton_utils import tl, triton
+
+from .index import prepare_chunk_indices
+from .utils import input_guard
+
+
+@triton.heuristics({'IS_VARLEN': lambda args: args['cu_seqlens'] is not None})
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps, num_stages=num_stages)
+        for num_warps in [1, 2, 4, 8] for num_stages in [2, 3, 4, 5]
+    ],
+    key=['BT'],
+)
+@triton.jit(do_not_specialize=['T'])
+def solve_tril_16x16_kernel(
+    A,
+    Ad,
+    cu_seqlens,
+    chunk_indices,
+    T,
+    H: tl.constexpr,
+    BT: tl.constexpr,
+    IS_VARLEN: tl.constexpr,
+):
+    i_t, i_bh = tl.program_id(0), tl.program_id(1)
+    i_b, i_h = i_bh // H, i_bh % H
+    if IS_VARLEN:
+        i_n, i_t = tl.load(chunk_indices + i_t * 2).to(
+            tl.int32), tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32)
+        bos, eos = tl.load(cu_seqlens + i_n).to(
+            tl.int32), tl.load(cu_seqlens + i_n + 1).to(tl.int32)
+        T = eos - bos
+    else:
+        bos, eos = i_b * T, i_b * T + T
+
+    A = A + (bos * H + i_h) * BT
+    Ad = Ad + (bos * H + i_h) * 16
+
+    offset = (i_t * 16) % BT
+    p_A = tl.make_block_ptr(A, (T, BT), (H * BT, 1), (i_t * 16, offset),
+                            (16, 16), (1, 0))
+    p_Ai = tl.make_block_ptr(Ad, (T, 16), (H * 16, 1), (i_t * 16, 0), (16, 16),
+                             (1, 0))
+    b_A = tl.load(p_A, boundary_check=(0, 1)).to(tl.float32)
+    b_A = -tl.where(
+        tl.arange(0, 16)[:, None] > tl.arange(0, 16)[None, :], b_A, 0)
+
+    o_i = tl.arange(0, 16)
+    for i in range(1, min(16, T - i_t * 16)):
+        b_a = -tl.load(A + (i_t * 16 + i) * H * BT + o_i + offset)
+        b_a = b_a + tl.sum(b_a[:, None] * b_A, 0)
+        mask = o_i == i
+        b_A = tl.where(mask[:, None], b_a, b_A)
+    b_A += o_i[:, None] == o_i[None, :]
+    tl.store(p_Ai,
+             b_A.to(p_Ai.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+
+
+@triton.heuristics({'IS_VARLEN': lambda args: args['cu_seqlens'] is not None})
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps, num_stages=num_stages)
+        for num_warps in [1, 2, 4, 8] for num_stages in [2, 3, 4, 5]
+    ],
+    key=['H', 'BT', 'IS_VARLEN'],
+)
+@triton.jit(do_not_specialize=['T'])
+def merge_16x16_to_32x32_inverse_kernel(A, Ad, Ai, cu_seqlens, chunk_indices,
+                                        T, H: tl.constexpr, BT: tl.constexpr,
+                                        IS_VARLEN: tl.constexpr):
+    i_t, i_bh = tl.program_id(0), tl.program_id(1)
+    i_b, i_h = i_bh // H, i_bh % H
+    if IS_VARLEN:
+        i_n, i_t = tl.load(chunk_indices + i_t * 2).to(
+            tl.int32), tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32)
+        bos, eos = tl.load(cu_seqlens + i_n).to(
+            tl.int32), tl.load(cu_seqlens + i_n + 1).to(tl.int32)
+        T = eos - bos
+    else:
+        bos, eos = i_b * T, i_b * T + T
+
+    A += (bos * H + i_h) * 32
+    Ad += (bos * H + i_h) * 16
+    Ai += (bos * H + i_h) * 32
+
+    p_A_21 = tl.make_block_ptr(A, (T, 32), (H * 32, 1), (i_t * 32 + 16, 0),
+                               (16, 16), (1, 0))
+    p_Ad_11 = tl.make_block_ptr(Ad, (T, 16), (H * 16, 1), (i_t * 32, 0),
+                                (16, 16), (1, 0))
+    p_Ad_22 = tl.make_block_ptr(Ad, (T, 16), (H * 16, 1), (i_t * 32 + 16, 0),
+                                (16, 16), (1, 0))
+    p_Ai_11 = tl.make_block_ptr(Ai, (T, 32), (H * 32, 1), (i_t * 32, 0),
+                                (16, 16), (1, 0))
+    p_Ai_22 = tl.make_block_ptr(Ai, (T, 32), (H * 32, 1), (i_t * 32 + 16, 16),
+                                (16, 16), (1, 0))
+    p_Ai_21 = tl.make_block_ptr(Ai, (T, 32), (H * 32, 1), (i_t * 32 + 16, 0),
+                                (16, 16), (1, 0))
+
+    A_21 = tl.load(p_A_21, boundary_check=(0, 1)).to(tl.float32)
+    Ai_11 = tl.load(p_Ad_11, boundary_check=(0, 1)).to(tl.float32)
+    Ai_22 = tl.load(p_Ad_22, boundary_check=(0, 1)).to(tl.float32)
+    Ai_21 = -tl.dot(tl.dot(Ai_22, A_21, input_precision='ieee'),
+                    Ai_11,
+                    input_precision='ieee')
+    tl.store(p_Ai_11,
+             Ai_11.to(p_Ai_11.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_22,
+             Ai_22.to(p_Ai_22.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_21,
+             Ai_21.to(p_Ai_21.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+
+
+@triton.heuristics({'IS_VARLEN': lambda args: args['cu_seqlens'] is not None})
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps, num_stages=num_stages)
+        for num_warps in [2, 4, 8] for num_stages in [2, 3, 4, 5]
+    ],
+    key=['H', 'BT', 'IS_VARLEN'],
+)
+@triton.jit(do_not_specialize=['T'])
+def merge_16x16_to_64x64_inverse_kernel(A, Ad, Ai, cu_seqlens, chunk_indices,
+                                        T, H: tl.constexpr, BT: tl.constexpr,
+                                        IS_VARLEN: tl.constexpr):
+    i_t, i_bh = tl.program_id(0), tl.program_id(1)
+    i_b, i_h = i_bh // H, i_bh % H
+    if IS_VARLEN:
+        i_n, i_t = tl.load(chunk_indices + i_t * 2).to(
+            tl.int32), tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32)
+        bos, eos = tl.load(cu_seqlens + i_n).to(
+            tl.int32), tl.load(cu_seqlens + i_n + 1).to(tl.int32)
+        T = eos - bos
+    else:
+        bos, eos = i_b * T, i_b * T + T
+
+    A += (bos * H + i_h) * 64
+    Ad += (bos * H + i_h) * 16
+    Ai += (bos * H + i_h) * 64
+
+    p_A_21 = tl.make_block_ptr(A, (T, 64), (H * 64, 1), (i_t * 64 + 16, 0),
+                               (16, 16), (1, 0))
+    p_A_32 = tl.make_block_ptr(A, (T, 64), (H * 64, 1), (i_t * 64 + 32, 16),
+                               (16, 16), (1, 0))
+    p_A_31 = tl.make_block_ptr(A, (T, 64), (H * 64, 1), (i_t * 64 + 32, 0),
+                               (16, 16), (1, 0))
+    p_A_43 = tl.make_block_ptr(A, (T, 64), (H * 64, 1), (i_t * 64 + 48, 32),
+                               (16, 16), (1, 0))
+    p_A_42 = tl.make_block_ptr(A, (T, 64), (H * 64, 1), (i_t * 64 + 48, 16),
+                               (16, 16), (1, 0))
+    p_A_41 = tl.make_block_ptr(A, (T, 64), (H * 64, 1), (i_t * 64 + 48, 0),
+                               (16, 16), (1, 0))
+    p_Ad_11 = tl.make_block_ptr(Ad, (T, 16), (H * 16, 1), (i_t * 64, 0),
+                                (16, 16), (1, 0))
+    p_Ad_22 = tl.make_block_ptr(Ad, (T, 16), (H * 16, 1), (i_t * 64 + 16, 0),
+                                (16, 16), (1, 0))
+    p_Ad_33 = tl.make_block_ptr(Ad, (T, 16), (H * 16, 1), (i_t * 64 + 32, 0),
+                                (16, 16), (1, 0))
+    p_Ad_44 = tl.make_block_ptr(Ad, (T, 16), (H * 16, 1), (i_t * 64 + 48, 0),
+                                (16, 16), (1, 0))
+
+    A_21 = tl.load(p_A_21, boundary_check=(0, 1)).to(tl.float32)
+    A_32 = tl.load(p_A_32, boundary_check=(0, 1)).to(tl.float32)
+    A_31 = tl.load(p_A_31, boundary_check=(0, 1)).to(tl.float32)
+    A_43 = tl.load(p_A_43, boundary_check=(0, 1)).to(tl.float32)
+    A_42 = tl.load(p_A_42, boundary_check=(0, 1)).to(tl.float32)
+    A_41 = tl.load(p_A_41, boundary_check=(0, 1)).to(tl.float32)
+
+    Ai_11 = tl.load(p_Ad_11, boundary_check=(0, 1)).to(tl.float32)
+    Ai_22 = tl.load(p_Ad_22, boundary_check=(0, 1)).to(tl.float32)
+    Ai_33 = tl.load(p_Ad_33, boundary_check=(0, 1)).to(tl.float32)
+    Ai_44 = tl.load(p_Ad_44, boundary_check=(0, 1)).to(tl.float32)
+
+    Ai_21 = -tl.dot(tl.dot(Ai_22, A_21, input_precision='ieee'),
+                    Ai_11,
+                    input_precision='ieee')
+    Ai_32 = -tl.dot(tl.dot(Ai_33, A_32, input_precision='ieee'),
+                    Ai_22,
+                    input_precision='ieee')
+    Ai_43 = -tl.dot(tl.dot(Ai_44, A_43, input_precision='ieee'),
+                    Ai_33,
+                    input_precision='ieee')
+
+    Ai_31 = -tl.dot(Ai_33,
+                    tl.dot(A_31, Ai_11, input_precision='ieee') +
+                    tl.dot(A_32, Ai_21, input_precision='ieee'),
+                    input_precision='ieee')
+    Ai_42 = -tl.dot(Ai_44,
+                    tl.dot(A_42, Ai_22, input_precision='ieee') +
+                    tl.dot(A_43, Ai_32, input_precision='ieee'),
+                    input_precision='ieee')
+    Ai_41 = -tl.dot(Ai_44,
+                    tl.dot(A_41, Ai_11, input_precision='ieee') +
+                    tl.dot(A_42, Ai_21, input_precision='ieee') +
+                    tl.dot(A_43, Ai_31, input_precision='ieee'),
+                    input_precision='ieee')
+
+    p_Ai_11 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64, 0),
+                                (16, 16), (1, 0))
+    p_Ai_22 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 16, 16),
+                                (16, 16), (1, 0))
+    p_Ai_33 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 32, 32),
+                                (16, 16), (1, 0))
+    p_Ai_44 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 48, 48),
+                                (16, 16), (1, 0))
+    p_Ai_21 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 16, 0),
+                                (16, 16), (1, 0))
+    p_Ai_31 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 32, 0),
+                                (16, 16), (1, 0))
+    p_Ai_32 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 32, 16),
+                                (16, 16), (1, 0))
+    p_Ai_41 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 48, 0),
+                                (16, 16), (1, 0))
+    p_Ai_42 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 48, 16),
+                                (16, 16), (1, 0))
+    p_Ai_43 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 48, 32),
+                                (16, 16), (1, 0))
+    tl.store(p_Ai_11,
+             Ai_11.to(p_Ai_11.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_22,
+             Ai_22.to(p_Ai_22.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_33,
+             Ai_33.to(p_Ai_33.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_44,
+             Ai_44.to(p_Ai_44.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_21,
+             Ai_21.to(p_Ai_21.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_31,
+             Ai_31.to(p_Ai_31.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_32,
+             Ai_32.to(p_Ai_32.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_41,
+             Ai_41.to(p_Ai_41.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_42,
+             Ai_42.to(p_Ai_42.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_43,
+             Ai_43.to(p_Ai_43.dtype.element_ty, fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+
+    fill_zeros = tl.zeros((16, 16), dtype=tl.float32)
+    p_Ai_12 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64, 16),
+                                (16, 16), (1, 0))
+    p_Ai_13 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64, 32),
+                                (16, 16), (1, 0))
+    p_Ai_14 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64, 48),
+                                (16, 16), (1, 0))
+    p_Ai_23 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 16, 32),
+                                (16, 16), (1, 0))
+    p_Ai_24 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 16, 48),
+                                (16, 16), (1, 0))
+    p_Ai_34 = tl.make_block_ptr(Ai, (T, 64), (H * 64, 1), (i_t * 64 + 32, 48),
+                                (16, 16), (1, 0))
+    tl.store(p_Ai_12,
+             fill_zeros.to(p_Ai_12.dtype.element_ty,
+                           fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_13,
+             fill_zeros.to(p_Ai_13.dtype.element_ty,
+                           fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_14,
+             fill_zeros.to(p_Ai_14.dtype.element_ty,
+                           fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_23,
+             fill_zeros.to(p_Ai_23.dtype.element_ty,
+                           fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_24,
+             fill_zeros.to(p_Ai_24.dtype.element_ty,
+                           fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+    tl.store(p_Ai_34,
+             fill_zeros.to(p_Ai_34.dtype.element_ty,
+                           fp_downcast_rounding="rtne"),
+             boundary_check=(0, 1))
+
+
+@input_guard
+def solve_tril(A: torch.Tensor,
+               cu_seqlens: Optional[torch.Tensor] = None,
+               output_dtype: torch.dtype = torch.float) -> torch.Tensor:
+    """
+    Compute the inverse of the lower triangular matrix
+    A should be strictly lower triangular, i.e., A.triu() == 0.
+
+    Args:
+        A (torch.Tensor):
+            [B, T, H, K]
+        cu_seqlens (torch.Tensor):
+            The cumulative sequence lengths of the input tensor.
+            Default: None.
+        output_dtype (torch.dtype):
+            The dtype of the output tensor. Default: `torch.float`
+
+    Returns:
+        (I + A)^-1 with the same shape as A
+    """
+    assert A.shape[-1] in [16, 32, 64]
+
+    B, T, H, BT = A.shape
+    Ad = torch.empty(B,
+                     T,
+                     H,
+                     16,
+                     device=A.device,
+                     dtype=torch.float if BT != 16 else output_dtype)
+
+    chunk_indices = prepare_chunk_indices(
+        cu_seqlens, 16) if cu_seqlens is not None else None
+    NT = len(chunk_indices) if cu_seqlens is not None else triton.cdiv(T, 16)
+    solve_tril_16x16_kernel[NT, B * H](
+        A=A,
+        Ad=Ad,
+        cu_seqlens=cu_seqlens,
+        chunk_indices=chunk_indices,
+        T=T,
+        H=H,
+        BT=BT,
+    )
+    if BT == 16:
+        return Ad
+
+    Ai = torch.empty(B, T, H, BT, device=A.device, dtype=output_dtype)
+    merge_fn = merge_16x16_to_32x32_inverse_kernel if BT == 32 else merge_16x16_to_64x64_inverse_kernel
+    chunk_indices = prepare_chunk_indices(
+        cu_seqlens, BT) if cu_seqlens is not None else None
+    NT = len(chunk_indices) if cu_seqlens is not None else triton.cdiv(T, BT)
+    merge_fn[NT, B * H](
+        A=A,
+        Ad=Ad,
+        Ai=Ai,
+        cu_seqlens=cu_seqlens,
+        chunk_indices=chunk_indices,
+        T=T,
+        H=H,
+        BT=BT,
+    )
+    return Ai

vllm/model_executor/layers/fla/ops/utils.py

@@ -0,0 +1,180 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
+#
+# This file contains code copied from the flash-linear-attention project.
+# The original source code was licensed under the MIT license and included
+# the following copyright notice:
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+# ruff: noqa: E501
+import contextlib
+import functools
+import logging
+import os
+from enum import Enum
+from typing import Any, Callable, Literal, Optional
+
+import torch
+
+from vllm.triton_utils import triton
+
+logger = logging.getLogger(__name__)
+
+COMPILER_MODE = os.getenv("FLA_COMPILER_MODE") == "1"
+FLA_CI_ENV = os.getenv("FLA_CI_ENV") == "1"
+FLA_GDN_FIX_BT = os.getenv("FLA_GDN_FIX_BT", "0") == "1"
+
+SUPPRESS_LEVEL = int(os.getenv("GDN_RECOMPUTE_SUPPRESS_LEVEL", "0"))
+
+
+def tensor_cache(
+        fn: Callable[..., torch.Tensor]) -> Callable[..., torch.Tensor]:
+    """
+    A decorator that caches the most recent results of a function with tensor inputs.
+
+    This decorator will store the output of the decorated function for the most recent set of input tensors.
+    The cache is limited to a fixed size (default is 4). When the cache is full, the oldest entry will be removed.
+
+    Args:
+        fn (Callable[..., torch.Tensor]):
+            The function to be decorated. It should take tensor inputs and return tensor outputs.
+
+    Returns:
+        Callable[..., torch.Tensor]:
+            A wrapped version of the input function with single-entry caching.
+    """
+
+    cache_entries: tuple[Optional[tuple], Optional[dict], Any] = []
+    cache_size = 4
+
+    @functools.wraps(fn)
+    def wrapper(*args: Any, **kwargs: Any) -> Any:
+        nonlocal cache_entries, cache_size
+        for i, entry in enumerate(cache_entries):
+            last_args, last_kwargs, last_result = entry
+            if len(args) == len(last_args) and len(kwargs) == len(last_kwargs) \
+                and all(a is b for a, b in zip(args, last_args)) \
+                and all(k in last_kwargs and v is last_kwargs[k] for k, v in kwargs.items()):
+                cache_entries = cache_entries[:i] + cache_entries[i + 1:] + [
+                    (args, kwargs, last_result)
+                ]
+                return last_result
+
+        result = fn(*args, **kwargs)
+
+        if len(cache_entries) >= cache_size:
+            cache_entries = cache_entries[1:]
+        cache_entries.append((args, kwargs, result))
+        return result
+
+    return wrapper
+
+
+def input_guard(
+        fn: Callable[..., torch.Tensor]) -> Callable[..., torch.Tensor]:
+    """
+    A decorator to make sure all input tensors are contiguous and set the device based on input tensors.
+    """
+
+    @functools.wraps(fn)
+    def wrapper(*args, **kwargs):
+        contiguous_args = (i if not isinstance(i, torch.Tensor) else
+                           i.contiguous() for i in args)
+        contiguous_kwargs = {
+            k: (v if not isinstance(v, torch.Tensor) else v.contiguous())
+            for k, v in kwargs.items()
+        }
+
+        tensor = None
+        for arg in args:
+            if isinstance(arg, torch.Tensor):
+                tensor = arg
+                break
+        if tensor is None:
+            for value in kwargs.values():
+                if isinstance(value, torch.Tensor):
+                    tensor = value
+                    break
+
+        if tensor is not None:
+            ctx = torch.cuda.device(tensor.device.index)
+        else:
+            ctx = contextlib.nullcontext()
+
+        with ctx:
+            return fn(*contiguous_args, **contiguous_kwargs)
+
+    return wrapper
+
+
+@functools.cache
+def get_available_device() -> str:
+    try:
+        return triton.runtime.driver.active.get_current_target().backend
+    except BaseException:
+        return 'cpu'
+
+
+@functools.cache
+def _check_platform() -> Literal['nvidia', 'amd', 'intel', 'musa']:
+    device = get_available_device()
+    mapping = {
+        "cuda": "nvidia",
+        "hip": "amd",
+        "xpu": "intel",
+    }
+    # return the mapped value, or the original if not found
+    return mapping.get(device, device)
+
+
+# For AMD GPUs, the triton backend is 'hip', while for Nvidia GPUs, the triton backend is 'cuda'.
+# However, the torch backend is 'cuda' for both Nvidia and AMD GPUs.
+# Therefore, we need to check the triton backend to determine the actual GPU vendor.
+device = get_available_device() if get_available_device() != 'hip' else 'cuda'
+device_torch_lib = getattr(torch, device)
+device_platform = _check_platform()
+
+is_amd = (device_platform == 'amd')
+is_intel = (device_platform == 'intel')
+is_nvidia = (device_platform == 'nvidia')
+is_intel_alchemist = (is_intel
+                      and 'Intel(R) Arc(TM) A' in torch.xpu.get_device_name(0))
+is_nvidia_hopper = (is_nvidia
+                    and ('NVIDIA H' in torch.cuda.get_device_name(0)
+                         or torch.cuda.get_device_capability()[0] >= 9))
+use_cuda_graph = (is_nvidia
+                  and os.environ.get('FLA_USE_CUDA_GRAPH', '0') == '1')
+
+
+def get_all_max_shared_mem():
+    try:
+        return [
+            triton.runtime.driver.active.utils.get_device_properties(i)
+            ['max_shared_mem'] for i in range(device_torch_lib.device_count())
+        ]
+    except BaseException:
+        return [-1]
+
+
+class Backend(Enum):
+    ADA = 101376  # RTX 4090
+    AMPERE = 166912  # A100
+    HOPPER = 232448  # H100
+    DEFAULT = 102400  # Default
+
+    @classmethod
+    def get_shared_memory(cls, arch: str) -> int:
+        try:
+            return cls[arch.upper()].value
+        except KeyError:
+            return cls.DEFAULT.value
+
+
+@functools.cache
+def check_shared_mem(arch: str = "none", tensor_idx: int = 0) -> bool:
+    try:
+        device_shared_mem_list = get_all_max_shared_mem()
+        max_shared_memory = device_shared_mem_list[tensor_idx]
+        return max_shared_memory >= Backend.get_shared_memory(arch)
+    except Exception:
+        return False

vllm/model_executor/layers/fla/ops/wy_fast.py

@@ -0,0 +1,114 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Songlin Yang, Yu Zhang
+#
+# This file contains code copied from the flash-linear-attention project.
+# The original source code was licensed under the MIT license and included
+# the following copyright notice:
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+# ruff: noqa: E501
+from typing import Optional
+
+import torch
+
+from vllm.triton_utils import tl, triton
+
+from .index import prepare_chunk_indices
+
+
+@triton.heuristics({'IS_VARLEN': lambda args: args['cu_seqlens'] is not None})
+@triton.autotune(
+    configs=[
+        triton.Config({}, num_warps=num_warps, num_stages=num_stages)
+        for num_warps in [2, 4, 8] for num_stages in [2, 3, 4]
+    ],
+    key=['H', 'K', 'V', 'BT', 'BK', 'BV', 'IS_VARLEN'],
+)
+@triton.jit(do_not_specialize=['T'])
+def recompute_w_u_fwd_kernel(k, v, beta, w, u, A, g, cu_seqlens, chunk_indices,
+                             T, H: tl.constexpr, Hg: tl.constexpr,
+                             K: tl.constexpr, V: tl.constexpr,
+                             BT: tl.constexpr, BK: tl.constexpr,
+                             BV: tl.constexpr, IS_VARLEN: tl.constexpr):
+    i_t, i_bh = tl.program_id(0), tl.program_id(1)
+    i_b, i_h = i_bh // H, i_bh % H
+    if IS_VARLEN:
+        i_n, i_t = tl.load(chunk_indices + i_t * 2).to(
+            tl.int32), tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32)
+        bos, eos = tl.load(cu_seqlens + i_n).to(
+            tl.int32), tl.load(cu_seqlens + i_n + 1).to(tl.int32)
+        T = eos - bos
+    else:
+        bos, eos = i_b * T, i_b * T + T
+    p_beta = tl.make_block_ptr(beta + bos * H + i_h, (T, ), (H, ),
+                               (i_t * BT, ), (BT, ), (0, ))
+    p_g = tl.make_block_ptr(g + (bos * H + i_h), (T, ), (H, ), (i_t * BT, ),
+                            (BT, ), (0, ))
+    p_A = tl.make_block_ptr(A + (bos * H + i_h) * BT, (T, BT), (H * BT, 1),
+                            (i_t * BT, 0), (BT, BT), (1, 0))
+    b_beta = tl.load(p_beta, boundary_check=(0, ))
+    b_A = tl.load(p_A, boundary_check=(0, 1))
+    b_g = tl.exp(tl.load(p_g, boundary_check=(0, )))
+
+    for i_v in range(tl.cdiv(V, BV)):
+        p_v = tl.make_block_ptr(v + (bos * H + i_h) * V, (T, V), (H * V, 1),
+                                (i_t * BT, i_v * BV), (BT, BV), (1, 0))
+        p_u = tl.make_block_ptr(u + (bos * H + i_h) * V, (T, V), (H * V, 1),
+                                (i_t * BT, i_v * BV), (BT, BV), (1, 0))
+        b_v = tl.load(p_v, boundary_check=(0, 1))
+        b_vb = (b_v * b_beta[:, None]).to(b_v.dtype)
+        b_u = tl.dot(b_A, b_vb, allow_tf32=False)
+        tl.store(p_u, b_u.to(p_u.dtype.element_ty), boundary_check=(0, 1))
+
+    for i_k in range(tl.cdiv(K, BK)):
+        p_k = tl.make_block_ptr(k + (bos * Hg + i_h // (H // Hg)) * K, (T, K),
+                                (Hg * K, 1), (i_t * BT, i_k * BK), (BT, BK),
+                                (1, 0))
+        p_w = tl.make_block_ptr(w + (bos * H + i_h) * K, (T, K), (H * K, 1),
+                                (i_t * BT, i_k * BK), (BT, BK), (1, 0))
+        b_k = tl.load(p_k, boundary_check=(0, 1))
+        b_kb = (b_k * b_beta[:, None] * b_g[:, None]).to(b_k.dtype)
+        b_w = tl.dot(b_A, b_kb)
+        tl.store(p_w, b_w.to(p_w.dtype.element_ty), boundary_check=(0, 1))
+
+
+def recompute_w_u_fwd(
+    k: torch.Tensor,
+    v: torch.Tensor,
+    beta: torch.Tensor,
+    g_cumsum: torch.Tensor,
+    A: torch.Tensor,
+    cu_seqlens: Optional[torch.LongTensor],
+) -> tuple[torch.Tensor, torch.Tensor]:
+    B, T, Hg, K, V = *k.shape, v.shape[-1]
+    H = v.shape[-2]
+    BT = A.shape[-1]
+
+    chunk_indices = prepare_chunk_indices(
+        cu_seqlens, BT) if cu_seqlens is not None else None
+    NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)
+    BK = 64
+    BV = 64
+    u = torch.empty_like(v)
+    w = k.new_empty(B, T, H, K)
+    recompute_w_u_fwd_kernel[(NT, B * H)](
+        k=k,
+        v=v,
+        beta=beta,
+        w=w,
+        u=u,
+        A=A,
+        g=g_cumsum,
+        cu_seqlens=cu_seqlens,
+        chunk_indices=chunk_indices,
+        T=T,
+        H=H,
+        Hg=Hg,
+        K=K,
+        V=V,
+        BT=BT,
+        BK=BK,
+        BV=BV,
+    )
+    return w, u