[DeepSeek-V4.1] Bump FlashMLA to the fork's rebase head (v4.1 kernels) (#39171)

Co-authored-by: Chunan Zeng <zcnrex@gmail.com>
This commit is contained in:
Baizhou Zhang
2026-09-13 15:43:19 -07:00
committed by GitHub
co-authored by Chunan Zeng
parent ff228d11fe
commit 5ebb16005d
3 changed files with 91 additions and 48 deletions
+57 -43
View File
@@ -1,9 +1,9 @@
# flash_mla
# sm90 dense decode HEAD_DIM_K=512 support (sgl-project/FlashMLA#9, merged).
# DeepSeek v4.1 kernels merged into the SGLang fork (sgl-project/FlashMLA@3e18517).
FetchContent_Declare(
repo-flashmla
URL https://${GITHUB_ARTIFACTORY}/sgl-project/FlashMLA/archive/c1dee569a494b184811a08171a690ece21420262.tar.gz
URL_HASH SHA256=77d3f1714b5903dc8f7a99fbc3a5d9a2b886e449f994b6c4b1d2feeacb467b1e
URL https://${GITHUB_ARTIFACTORY}/sgl-project/FlashMLA/archive/3e18517fb055a6c9608eef5a1f1347fb1a047bbd.tar.gz
URL_HASH SHA256=ab2af4657683a1bbaa707a2781a5e2c2792eba574ff36905f1832598e67fea79
)
FetchContent_Populate(repo-flashmla)
@@ -42,23 +42,16 @@ if(${CUDA_VERSION} VERSION_GREATER 12.8)
set(FLASHMLA_ENABLE_SM100 ON)
endif()
if(${CUDA_VERSION} VERSION_GREATER_EQUAL "13.0")
# Patch FlashMLA sources for SM103a support.
# These patches are only needed (and only valid) with CUDA 13+.
# Patch utils.h: widen IS_SM100 to cover the full SM100 family.
# Newer FlashMLA versions use csrc/utils.h.
set(FLASHMLA_UTILS_FILE "${repo-flashmla_SOURCE_DIR}/csrc/utils.h")
file(READ "${FLASHMLA_UTILS_FILE}" FLASHMLA_UTILS_CONTENT)
string(REPLACE
"#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ == 1000)
#define IS_SM100 1"
"#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000) && (__CUDA_ARCH__ < 1100)
#define IS_SM100 1"
FLASHMLA_UTILS_CONTENT "${FLASHMLA_UTILS_CONTENT}")
file(WRITE "${FLASHMLA_UTILS_FILE}" "${FLASHMLA_UTILS_CONTENT}")
message(STATUS "Patched utils.h for SM103a support")
# B300 is sm_103; sm_100f is family-compatible so it runs there, but ptxas emits
# better SASS for the native arch. FlashMLA's setup.py picks sm_100a + sm_103a
# over sm_100f for the same reason. The cutlass patch below is what lets a TU
# compile at __CUDA_ARCH__ == 1030 against the pinned cutlass.
list(APPEND FLASHMLA_CUDA_FLAGS
"-gencode=arch=compute_103a,code=sm_103a"
)
# Patch cutlass/arch/config.h: add SM103 architecture defines.
# This patch is only needed (and only valid) with CUDA 13+.
# The new block is inserted right before the existing "// SM101 and SM101a"
# anchor in the upstream header.
set(CUTLASS_CONFIG_FILE "${repo-flashmla_SOURCE_DIR}/csrc/cutlass/include/cutlass/arch/config.h")
@@ -99,26 +92,33 @@ set(FlashMLA_SOURCES
# Compatibility shim for sgl-kernel torch.ops API.
${repo-flashmla_SOURCE_DIR}/csrc/python_api.cpp
# Attention entry points (the pybind registrations in these files are
# compiled out by FLASH_MLA_LIBTORCH_ONLY).
${repo-flashmla_SOURCE_DIR}/csrc/api/dense_decode.cpp
${repo-flashmla_SOURCE_DIR}/csrc/api/sparse_decode.cpp
${repo-flashmla_SOURCE_DIR}/csrc/api/sparse_prefill.cpp
# Decode metadata/combine kernels.
${repo-flashmla_SOURCE_DIR}/csrc/smxx/decode/get_decoding_sched_meta/get_decoding_sched_meta.cu
${repo-flashmla_SOURCE_DIR}/csrc/smxx/decode/combine/combine.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/smxx/decode/get_decoding_sched_meta/get_decoding_sched_meta.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/smxx/decode/combine/combine.cu
# sm90 dense decode.
${repo-flashmla_SOURCE_DIR}/csrc/sm90/decode/dense/instantiations/fp16.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm90/decode/dense/instantiations/bf16.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/decode/dense/instantiations/fp16.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/decode/dense/instantiations/bf16.cu
# sm90 sparse decode.
${repo-flashmla_SOURCE_DIR}/csrc/sm90/decode/sparse_fp8/instantiations/model1_persistent_h64.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm90/decode/sparse_fp8/instantiations/model1_persistent_h128.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm90/decode/sparse_fp8/instantiations/v32_persistent_h64.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm90/decode/sparse_fp8/instantiations/v32_persistent_h128.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/decode/sparse/instantiations/v4_persistent_h64.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/decode/sparse/instantiations/v4_persistent_h128.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/decode/sparse/instantiations/v32_persistent_h64.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/decode/sparse/instantiations/v32_persistent_h128.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/decode/sparse/instantiations/v32_no_rope_persistent_h64.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/decode/sparse/instantiations/v32_no_rope_persistent_h128.cu
# sm90 sparse prefill.
${repo-flashmla_SOURCE_DIR}/csrc/sm90/prefill/sparse/fwd.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm90/prefill/sparse/instantiations/phase1_k512.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm90/prefill/sparse/instantiations/phase1_k512_topklen.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm90/prefill/sparse/instantiations/phase1_k576.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm90/prefill/sparse/instantiations/phase1_k576_topklen.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/prefill/sparse/instantiations/phase1_k512.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/prefill/sparse/instantiations/phase1_k512_topklen.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/prefill/sparse/instantiations/phase1_k576.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm90/prefill/sparse/instantiations/phase1_k576_topklen.cu
${repo-flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/dense_fp8_python_api.cpp
${repo-flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/flash_fwd_mla_fp8_sm90.cu
@@ -128,20 +128,33 @@ set(FlashMLA_SOURCES
if(FLASHMLA_ENABLE_SM100)
list(APPEND FlashMLA_SOURCES
# sm100 dense prefill/bwd.
${repo-flashmla_SOURCE_DIR}/csrc/sm100/prefill/dense/fmha_cutlass_fwd_sm100.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm100/prefill/dense/fmha_cutlass_bwd_sm100.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/dense/fmha_cutlass_fwd_sm100.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/dense/fmha_cutlass_bwd_sm100.cu
# sm100 sparse prefill.
${repo-flashmla_SOURCE_DIR}/csrc/sm100/prefill/sparse/fwd/head64/instantiations/phase1_k512.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm100/prefill/sparse/fwd/head64/instantiations/phase1_k576.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm100/prefill/sparse/fwd/head128/instantiations/phase1_k512.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm100/prefill/sparse/fwd/head128/instantiations/phase1_k576.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm100/prefill/sparse/fwd_for_small_topk/head128/instantiations/phase1_prefill_k512.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd/head64/instantiations/phase1_h64_k512.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd/head64/instantiations/phase1_h64_k576.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd/head128/instantiations/phase1_k512.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd/head128/instantiations/phase1_k576.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd_for_small_topk/head128/instantiations/phase1_k512.cu
# sm100 sparse decode.
${repo-flashmla_SOURCE_DIR}/csrc/sm100/decode/head64/instantiations/v32.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm100/decode/head64/instantiations/model1.cu
${repo-flashmla_SOURCE_DIR}/csrc/sm100/prefill/sparse/fwd_for_small_topk/head128/instantiations/phase1_decode_k512.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/decode/sparse/head64/instantiations/v32_h64.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/decode/sparse/head64/instantiations/v32_h64_no_split.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/decode/sparse/head64/instantiations/v32_no_rope_h64.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/decode/sparse/head64/instantiations/v32_no_rope_h64_no_split.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/decode/sparse/head64/instantiations/v4_h64.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/decode/sparse/head64/instantiations/v4_h64_no_split.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/decode/sparse/head64/instantiations/v41_h64.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/decode/sparse/head64/instantiations/v41_h64_no_split.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/decode/sparse/head64/instantiations/v41fp4_h64.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/decode/sparse/head64/instantiations/v41fp4_h64_no_split.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd_for_small_topk/head128/instantiations/phase1_decode_k512.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd_for_small_topk/head128/instantiations/phase1_decode_k512_splitkv.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd_for_small_topk/head128/instantiations/phase1_decode_k512_v41.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd_for_small_topk/head128/instantiations/phase1_decode_k512_v41_splitkv.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd_for_small_topk/head128/instantiations/phase1_decode_k512_v41fp4.cu
${repo-flashmla_SOURCE_DIR}/csrc/kernels/sm100/prefill/sparse/fwd_for_small_topk/head128/instantiations/phase1_decode_k512_v41fp4_splitkv.cu
)
endif()
@@ -167,7 +180,6 @@ endif()
target_include_directories(flashmla_ops PRIVATE
${repo-flashmla_SOURCE_DIR}/csrc
${repo-flashmla_SOURCE_DIR}/csrc/kerutils/include
${repo-flashmla_SOURCE_DIR}/csrc/sm90
${repo-flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/
${repo-flashmla_SOURCE_DIR}/csrc/cutlass/include
${repo-flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
@@ -178,4 +190,6 @@ target_link_libraries(flashmla_ops PRIVATE ${TORCH_LIBRARIES} c10 cuda)
install(TARGETS flashmla_ops LIBRARY DESTINATION "sgl_kernel")
target_compile_definitions(flashmla_ops PRIVATE)
# FlashMLA's csrc/api/*.cpp register their ops through pybind by default; sgl-kernel
# links them as plain libtorch C++ and registers the ops itself in flashmla_extension.cc.
target_compile_definitions(flashmla_ops PRIVATE FLASH_MLA_LIBTORCH_ONLY)
@@ -16,11 +16,36 @@ limitations under the License.
#include <torch/all.h>
#include <torch/library.h>
#include "api/dense_decode.h"
#include "api/sparse_decode.h"
#include "api/sparse_fwd.h"
#include "sgl_kernel_ops.h"
// FlashMLA exposes these two entry points only from csrc/api/*.cpp (its own pybind layer lives
// behind FLASH_MLA_LIBTORCH_ONLY), so declare them here the way csrc/python_api.cpp does.
std::tuple<at::Tensor, at::Tensor, std::optional<at::Tensor>, std::optional<at::Tensor>> dense_attn_decode_interface(
at::Tensor& q,
const at::Tensor& kcache,
const int head_size_v,
const at::Tensor& seqlens_k,
const at::Tensor& block_table,
const float softmax_scale,
bool is_causal,
std::optional<at::Tensor>& tile_scheduler_metadata,
std::optional<at::Tensor>& num_splits);
std::tuple<at::Tensor, at::Tensor, std::optional<at::Tensor>, std::optional<at::Tensor>> sparse_attn_decode_interface(
const at::Tensor& q,
const at::Tensor& kv,
const at::Tensor& indices,
const std::optional<at::Tensor>& topk_length,
const std::optional<at::Tensor>& attn_sink,
std::optional<at::Tensor>& tile_scheduler_metadata,
std::optional<at::Tensor>& num_splits,
const std::optional<at::Tensor>& extra_kv,
const std::optional<at::Tensor>& extra_indices,
const std::optional<at::Tensor>& extra_topk_length,
int d_v,
float sm_scale,
const std::optional<std::string>& kv_format);
static std::tuple<at::Tensor, at::Tensor, std::optional<at::Tensor>, std::optional<at::Tensor>> sgl_sparse_decode_fwd(
const at::Tensor& q,
const at::Tensor& kv,
@@ -46,7 +71,8 @@ static std::tuple<at::Tensor, at::Tensor, std::optional<at::Tensor>, std::option
extra_indices,
extra_topk_length,
static_cast<int>(d_v),
static_cast<float>(sm_scale));
static_cast<float>(sm_scale),
std::nullopt);
}
static std::tuple<at::Tensor, at::Tensor, std::optional<at::Tensor>, std::optional<at::Tensor>> sgl_dense_decode_fwd(
@@ -58,8 +58,11 @@ class BasicDecodeCorrectnessMixin:
# Language-agnostic gibberish detector. Healthy English output is
# >90% printable ASCII; multilingual token salad / Unicode noise
# from broken weight load drops well below 50%.
# Q/A framing, as in the probes above: a base model continues a bare
# instruction with arbitrary text whose language is not pinned down,
# so the ratio would measure the continuation, not output health.
out = self._decode_generate(
"Write a single sentence about a sunny day in the park.",
"Q: Write a single sentence about a sunny day in the park.\nA:",
self.sanity_max_new_tokens_long,
)
printable = sum(1 for c in out if 32 <= ord(c) < 127 or c in "\n\t")