[DSv4] Adding TRTLLM gen attention kernel (#43827)

This commit is contained in:
Yongye Zhu
2026-06-04 07:35:09 -07:00
committed by GitHub
parent 3e77036768
commit b5235fca2e
20 changed files with 2659 additions and 86 deletions
+14
View File
@@ -228,3 +228,17 @@ MLA decode backends are selected using the standard
| `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 |
| `XPU_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | Any | 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | Any |
### DeepSeek V4 Decode Backends
DeepSeek V4 sparse MLA uses its own decode backends, selected via
`--attention-backend=<BACKEND>` (e.g., `FLASHMLA_SPARSE_DSV4`,
`FLASHINFER_MLA_SPARSE_DSV4`). They share the V4 sparse-index
pipeline (compressor + SWA + indexer, 256-token blocks, head 512);
default on NVIDIA is `FLASHMLA_SPARSE_DSV4`.
| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
| ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | ------ | --------- | --- | --------------- | ------------ |
| `FLASHINFER_MLA_SPARSE_DSV4` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | Any |
| `FLASHMLA_SPARSE_DSV4` | fp16, bf16 | `auto` | 256 | 512 | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | Any |
| `ROCM_FLASHMLA_SPARSE_DSV4` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |