Skip to content

[TRTLLM-12807][feat] add CuteDSL FP8/FP16 MLA decode attention fmha lib #17266

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

Closed
brnguyen2 wants to merge 29 commits into from
Closed

[TRTLLM-12807][feat] add CuteDSL FP8/FP16 MLA decode attention fmha lib #17266

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
29 commits
Select commit Hold shift + click to select a range
e37ef7f
[None][feat] add CuteDSL FP8/FP16 MLA decode attention backend
haow-nv Jun 8, 2026
4c01205
[None][feat] CuteDSL MLA decode: fp8 scaling fix, BF16 support, per-l…
haow-nv Jun 12, 2026
7678862
[None][fix] CuteDSL MLA decode: v1 paged-KV layout + CUDA-graph-safe …
haow-nv Jun 16, 2026
748ffeb
[TRTLLM-12807][feat] Wrap CuteDSL MLA decode as FMHA lib
yuxianq Jun 13, 2026
66159f1
[None][feat] CuteDSL MLA decode: MTP causal mask + can_implement gate
haow-nv Jun 24, 2026
0fe7413
[None][feat] CuteDSL MLA decode: persistent + split-KV alignment (fla…
haow-nv Jun 29, 2026
7aed06f
[None][feat] CuteDSL MLA decode: mixed prefill+decode page-table offs…
haow-nv Jun 30, 2026
cb20915
[None][feat] CuteDSL MLA decode: autotuned is_persistent tactic, perf…
haow-nv Jul 6, 2026
0960936
[None][chore] CuteDSL MLA: drop stale is_available comment
haow-nv Jul 6, 2026
722a7de
[None][chore] CuteDSL MLA decode: comment cleanup, drop default_is_pe…
haow-nv Jul 6, 2026
e396dc1
[None][chore] CuteDSL MLA decode: apply pre-commit formatting, fix co…
haow-nv Jul 6, 2026
5d6bc50
[None][chore] CuteDSL MLA decode: drop standalone decode unit test
haow-nv Jul 7, 2026
9bc6bc1
[None][fix] CuteDSL MLA: reject sparse attention in FMHA gate
haow-nv Jul 7, 2026
b082608
[None][chore] CuteDSL MLA decode: trim comments, drop unused runner i…
haow-nv Jul 7, 2026
cb08644
[None][perf] CuteDSL MLA: gate (128,1) decode on spec-decode off
haow-nv Jul 7, 2026
3968be1
[None][perf] CuteDSL MLA: batch-aware perf gate, autotuner decode war…
haow-nv Jul 10, 2026
de72d57
[None][chore] CuteDSL MLA: apply pre-commit formatting
haow-nv Jul 10, 2026
d0949c6
[None][perf] CuteDSL MLA: restructure decode perf gate, drop (128,1) …
haow-nv Jul 14, 2026
66b8e90
[None][fix] stabilize CuteDSL MLA split-KV workspace for CUDA graphs
haow-nv Jul 15, 2026
a9d21f5
[None][chore] CuteDSL MLA decode: drop debug kernel-arg dumps, fix au…
haow-nv Jul 23, 2026
60b0425
[None][fix] CuteDSL MLA decode: reject Helix, slice AutoTuner workspa…
haow-nv Jul 28, 2026
2915a9e
[None][fix] CuteDSL MLA decode: tighten split_kv and batch gating
haow-nv Jul 29, 2026
a575f97
[None][chore] CuteDSL MLA decode: rename fmha/cute_dsl.py to cute_dsl…
haow-nv Jul 29, 2026
faf5aba
[None][fix] CuteDSL MLA decode: keep dtype/heads gate during autotuning
haow-nv Aug 3, 2026
346fb8c
[None][fix] CuteDSL MLA decode: drop (128,4)/(128,8) from fp8 perf al…
haow-nv Aug 4, 2026
827cc9e
[None][fix] CuteDSL MLA decode: bucket the fallback tactic's batch size
brnguyen2 Aug 4, 2026
9f9e319
[None][test] CuteDSL MLA decode: cover the AutoTuner tuning path
brnguyen2 Aug 4, 2026
ee0319b
[None][test] Add disagg decode-only smoke for the CuTe DSL MLA FMHA lib
brnguyen2 Aug 4, 2026
780fa9b
[None][chore] Rename CuteDslMlaFmha FMHA lib to CuteDslFmha
brnguyen2 Aug 4, 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
6 changes: 6 additions & 0 deletions .pre-commit-config.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -140,6 +140,9 @@ common-files: &common_files |
tensorrt_llm/_torch/custom_ops/userbuffers_custom_ops.py |
tensorrt_llm/_torch/cute_dsl_kernels/__init__.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/__init__.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_decode_fp16.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_decode_fp8.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_helpers.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/custom_pipeline.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/dense_blockscaled_gemm_persistent.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/utils.py |
Expand Down Expand Up @@ -911,6 +914,9 @@ legacy-files: &legacy_files |
tensorrt_llm/_torch/custom_ops/userbuffers_custom_ops.py |
tensorrt_llm/_torch/cute_dsl_kernels/__init__.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/__init__.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_decode_fp16.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_decode_fp8.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_helpers.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/custom_pipeline.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/dense_blockscaled_gemm_persistent.py |
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/utils.py |
Expand Down
3 changes: 3 additions & 0 deletions legacy-files.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,6 +132,9 @@ tensorrt_llm/_torch/custom_ops/trtllm_gen_custom_ops.py
tensorrt_llm/_torch/custom_ops/userbuffers_custom_ops.py
tensorrt_llm/_torch/cute_dsl_kernels/__init__.py
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/__init__.py
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_decode_fp16.py
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_decode_fp8.py
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_helpers.py
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/custom_pipeline.py
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/dense_blockscaled_gemm_persistent.py
tensorrt_llm/_torch/cute_dsl_kernels/blackwell/utils.py
Expand Down
3 changes: 3 additions & 0 deletions pyproject.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -189,6 +189,9 @@ exclude = [
"tensorrt_llm/_torch/custom_ops/userbuffers_custom_ops.py",
"tensorrt_llm/_torch/cute_dsl_kernels/__init__.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/__init__.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_decode_fp16.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_decode_fp8.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_helpers.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/custom_pipeline.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/dense_blockscaled_gemm_persistent.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/utils.py",
Expand Down
3 changes: 3 additions & 0 deletions ruff-legacy.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -149,6 +149,9 @@ include = [
"tensorrt_llm/_torch/custom_ops/userbuffers_custom_ops.py",
"tensorrt_llm/_torch/cute_dsl_kernels/__init__.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/__init__.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_decode_fp16.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_decode_fp8.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/attention/mla/mla_helpers.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/custom_pipeline.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/dense_blockscaled_gemm_persistent.py",
"tensorrt_llm/_torch/cute_dsl_kernels/blackwell/utils.py",
Expand Down
2 changes: 2 additions & 0 deletions tensorrt_llm/_torch/attention_backend/fmha/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.

from .cute_dsl import CuteDslFmha
from .fallback import FallbackFmha
from .flashinfer_trtllm_gen import FlashInferTrtllmGenFmha
from .interface import Fmha
Expand All @@ -23,6 +24,7 @@
__all__ = [
"DEFAULT_FMHA_LIBS",
"FMHA_LIBS",
"CuteDslFmha",
"FallbackFmha",
"FlashInferTrtllmGenFmha",
"Fmha",
Expand Down
Loading
Loading