Skip to content

[Attention] TRITON_MLA_SPARSE backend for SM80/SM121 sparse MLA (rebase & takeover of #38476) #47629

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
thomaslwang wants to merge 10 commits into
base: main
Choose a base branch
from
+1,808 −9
Open

[Attention] TRITON_MLA_SPARSE backend for SM80/SM121 sparse MLA (rebase & takeover of #38476) #47629

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
fb50482
Add TRITON_MLA_SPARSE backend for SM80/SM121 sparse MLA support
haosdent May 14, 2026
8c8877b
Gate fused_indexer_q_rope_quant on SM89+ (Triton fp8e4nv store)
ngila-ai Jul 5, 2026
85a3c30
Add TRITON_MLA_SPARSE as BF16-KV sparse fallback on SM12x
ngila-ai Jul 5, 2026
a294a69
Declare MultipleOf(64) block sizes for TRITON_MLA_SPARSE
ngila-ai Jul 5, 2026
6a4ef95
Regenerate attention backend docs
ngila-ai Jul 5, 2026
7d3f3ac
Fix paged MQA logits tail store
ngila-ai Jul 22, 2026
f5739d6
Merge branch 'main' into triton-mla-sparse-sm80
thomaslwang Jul 23, 2026
94983be
Merge branch 'main' into triton-mla-sparse-sm80
thomaslwang Jul 31, 2026
5201cdf
Fix int32 overflow in paged MQA logits block addressing
ngila-ai Jul 31, 2026
064801d
Merge branch 'main' into triton-mla-sparse-sm80
thomaslwang Aug 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
34 changes: 34 additions & 0 deletions csrc/libtorch_stable/quantization/fp4/nvfp4_quant_entry.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,3 +173,37 @@ void silu_and_mul_scaled_fp4_experts_quant(
STD_TORCH_CHECK_NOT_IMPLEMENTED(
false, "No compiled silu_and_mul nvfp4 experts quantization kernel");
}

// MXFP4 link stubs for archs without a compiled MXFP4 kernel. On SM10.x the
// real impls in `mxfp4_experts_quant.cu` are compiled (same CMake block that
// defines `ENABLE_NVFP4_SM100`); on SM 8.x the real file is skipped, leaving
// `torch_bindings.cpp`'s unconditional `&mxfp4_experts_quant` reference with
// no symbol to bind to, so the `.so` fails to load. These stubs satisfy the
// linker on unsupported archs and raise NOT_IMPLEMENTED if anyone actually
// calls them.
#if !(defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100)
void mxfp4_experts_quant(
torch::stable::Tensor& output, torch::stable::Tensor& output_scale,
torch::stable::Tensor const& input,
torch::stable::Tensor const& input_offset_by_experts,
torch::stable::Tensor const& output_scale_offset_by_experts,
int64_t n_experts) {
STD_TORCH_CHECK_NOT_IMPLEMENTED(
false, "No compiled mxfp4 experts quantization kernel for SM ",
get_sm_version_num(),
". Recompile with SM 10.x (Blackwell) for MXFP4 support.");
}

void silu_and_mul_mxfp4_experts_quant(
torch::stable::Tensor& output, torch::stable::Tensor& output_scale,
torch::stable::Tensor const& input,
torch::stable::Tensor const& input_offset_by_experts,
torch::stable::Tensor const& output_scale_offset_by_experts,
int64_t n_experts) {
STD_TORCH_CHECK_NOT_IMPLEMENTED(
false,
"No compiled silu_and_mul mxfp4 experts quantization kernel for SM ",
get_sm_version_num(),
". Recompile with SM 10.x (Blackwell) for MXFP4 support.");
}
#endif // !ENABLE_NVFP4_SM100
17 changes: 17 additions & 0 deletions docs/design/attention_backends.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,6 +162,23 @@ hardware and configuration.
MLA decode backends are selected using the standard
`-ac.backend=<BACKEND>` argument (e.g., `FLASHMLA`, `TRITON_MLA`).

| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
| ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | ------ | --------- | --- | --------------- | ------------ |
| `CUTLASS_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 128 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
| `FLASHINFER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
| `FLASHINFER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
| `FLASHINFER_MLA_SPARSE_SM120` | bf16 | `auto`, `fp8`, `fp8_e4m3`, `fp8_ds_mla` | 64, 256 | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | 12.x |
| `FLASHMLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x-10.x |
| `FLASHMLA_SPARSE` | bf16 | `auto`, `bfloat16`, `fp8_ds_mla` | 64 | 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
| `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
| `FLASH_ATTN_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | 64 | Any | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x |
| `ROCM_AITER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %1 | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 1, 64 | Any | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
| `ROCM_AITER_TRITON_MLA` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
| `TOKENSPEED_MLA` | fp16, bf16 | `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
| `TRITON_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
| `TRITON_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | %64 | Any | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | Any |
| `XPU_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | Any | 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | Any |
--8<-- "gen:table-mla-decode"

### DeepSeek V4 Decode Backends
Expand Down
Loading
Loading