[Attention][DSA] Enable W4A16 DSA (#51724)

Signed-off-by: Shiyang Chen <shiychen@nvidia.com>
This commit is contained in:
sychen52
2026-09-01 12:37:47 +08:00
committed by GitHub
parent ce2e343be1
commit 7c5dc571cb
17 changed files with 824 additions and 22 deletions
+104
View File
@@ -7,6 +7,10 @@
#include "quantization/vectorization_utils.cuh"
#include "concat_mla_q.cuh"
#if !defined(USE_ROCM) && defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100
#include "nvfp4_ds_mla_cache.h"
#endif
#ifdef USE_ROCM
#include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
#else
@@ -917,6 +921,7 @@ void concat_and_cache_mla(
int pe_dim = k_pe.size(1);
int block_size = kv_cache.size(1);
const bool is_nvfp4_ds_mla = kv_cache_dtype == "nvfp4_ds_mla";
if (kv_cache_dtype == "fp8_ds_mla") {
STD_TORCH_CHECK(kv_lora_rank == 512,
"kv_lora_rank must be 512 for fp8_ds_mla");
@@ -927,6 +932,21 @@ void concat_and_cache_mla(
"kv_c.element_size() must be 2 for fp8_ds_mla");
STD_TORCH_CHECK(k_pe.element_size() == 2,
"k_pe.element_size() must be 2 for fp8_ds_mla");
} else if (is_nvfp4_ds_mla) {
constexpr int bytes_per_token = 352;
STD_TORCH_CHECK(kv_lora_rank == 512, "kv_lora_rank must be 512 for ",
kv_cache_dtype);
STD_TORCH_CHECK(pe_dim == 64, "pe_dim must be 64 for ", kv_cache_dtype);
STD_TORCH_CHECK(
kv_cache.size(2) == bytes_per_token / kv_cache.element_size(),
"kv_cache.size(2) must be ", bytes_per_token, " bytes for ",
kv_cache_dtype);
STD_TORCH_CHECK(
kv_c.scalar_type() == torch::headeronly::ScalarType::BFloat16,
"kv_c must be bfloat16 for ", kv_cache_dtype);
STD_TORCH_CHECK(
k_pe.scalar_type() == torch::headeronly::ScalarType::BFloat16,
"k_pe must be bfloat16 for ", kv_cache_dtype);
} else {
STD_TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
}
@@ -950,6 +970,24 @@ void concat_and_cache_mla(
dim3 block(96);
DISPATCH_BY_KV_CACHE_DTYPE(kv_c.scalar_type(), kv_cache_dtype,
CALL_CONCAT_AND_CACHE_DS_MLA);
} else if (is_nvfp4_ds_mla) {
// The kernels live in nvfp4_ds_mla_cache_kernels.cu, which is only built
// when the target list contains an SM100 arch.
#if !defined(USE_ROCM) && defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100
const cudaDeviceProp* props = get_device_prop();
STD_TORCH_CHECK(props->major == 10,
"nvfp4_ds_mla KV cache requires SM100 (Blackwell); "
"got SM",
props->major, props->minor);
vllm::launch_concat_and_cache_nvfp4_ds_mla(
kv_c.data_ptr(), k_pe.data_ptr(), kv_cache.data_ptr(),
slot_mapping.const_data_ptr<int64_t>(), block_stride, entry_stride,
kv_c_stride, k_pe_stride, block_size, num_tokens, stream);
#else
STD_TORCH_CHECK(false,
"the nvfp4 ds_mla kv-cache format requires a CUDA build "
"targeting SM100 (Blackwell)");
#endif
} else {
dim3 grid(num_tokens);
dim3 block(std::min(kv_lora_rank, 512));
@@ -1659,6 +1697,72 @@ void cp_gather_and_upconvert_fp8_kv_cache(
dst_entry_stride, seq_starts_ptr);
}
void cp_gather_and_upconvert_nvfp4_kv_cache(
torch::stable::Tensor const& src_cache, // [NUM_BLOCKS, BLOCK_SIZE,
// 352]
torch::stable::Tensor const& dst, // [TOT_TOKENS, 576]
torch::stable::Tensor const& block_table, // [BATCH, BLOCK_INDICES]
torch::stable::Tensor const& workspace_starts, // [BATCH]
int64_t batch_size) {
// The kernel lives in nvfp4_ds_mla_cache_kernels.cu, which is only built when
// the target list contains an SM100 arch.
#if !defined(USE_ROCM) && defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100
torch::stable::accelerator::DeviceGuard device_guard(
src_cache.get_device_index());
const cudaStream_t stream = get_current_cuda_stream();
const cudaDeviceProp* props = get_device_prop();
STD_TORCH_CHECK(props->major == 10,
"nvfp4_ds_mla KV cache requires SM100 (Blackwell); "
"got SM",
props->major, props->minor);
int32_t block_size = src_cache.size(1);
int32_t head_dim = dst.size(1);
STD_TORCH_CHECK(
block_table.scalar_type() == torch::headeronly::ScalarType::Int,
"block_table must be int32");
STD_TORCH_CHECK(
workspace_starts.scalar_type() == torch::headeronly::ScalarType::Int,
"workspace_starts must be int32");
STD_TORCH_CHECK(src_cache.device() == dst.device(),
"src_cache and dst must be on the same device");
STD_TORCH_CHECK(src_cache.device() == block_table.device(),
"src_cache and block_table must be on the same device");
STD_TORCH_CHECK(src_cache.device() == workspace_starts.device(),
"src_cache and workspace_starts must be on the same device");
STD_TORCH_CHECK(
src_cache.scalar_type() == torch::headeronly::ScalarType::Byte,
"src_cache must be uint8 for the nvfp4 ds_mla format, but got ",
src_cache.scalar_type());
STD_TORCH_CHECK(dst.scalar_type() == torch::headeronly::ScalarType::BFloat16,
"dst must be bfloat16");
STD_TORCH_CHECK(head_dim == 576, "head_dim must be 576 for MLA");
STD_TORCH_CHECK(src_cache.size(2) == 352,
"src_cache.size(2) must be 352 for the nvfp4 ds_mla format");
int64_t block_table_stride = block_table.stride(0);
int64_t cache_block_stride = src_cache.stride(0);
int64_t cache_entry_stride = src_cache.stride(1);
int64_t dst_entry_stride = dst.stride(0);
const int total_tokens = dst.size(0);
vllm::launch_cp_gather_and_upconvert_nvfp4_kv_cache(
src_cache.const_data_ptr<uint8_t>(), dst.data_ptr(),
block_table.const_data_ptr<int32_t>(),
workspace_starts.const_data_ptr<int32_t>(),
static_cast<int32_t>(batch_size), block_size, total_tokens,
block_table_stride, cache_block_stride, cache_entry_stride,
dst_entry_stride, stream);
#else
STD_TORCH_CHECK(false,
"the nvfp4 ds_mla kv-cache format requires a CUDA build "
"targeting SM100 (Blackwell)");
#endif
}
// Macro to dispatch the kernel based on the data type.
#define CALL_INDEXER_K_QUANT_AND_CACHE(KV_T, CACHE_T, KV_DTYPE) \
vllm::indexer_k_quant_and_cache_kernel<KV_T, CACHE_T, KV_DTYPE> \