Add nvfp4 attention support for vLLM serving#1898

Merged

kaix-nv merged 24 commits into

mainfrom

kaix/sparse_attn_quant_compact

Jul 16, 2026

Commits on Jul 15, 2026

Add compact vLLM attention quant carrier
kaix-nv
committed
Add compact NVFP4 V attention path
kaix-nv
committed
Add compact NVFP4 vLLM attention worker
kaix-nv
committed
Document compact NVFP4 vLLM attention serving
kaix-nv
committed
Harden compact NVFP4 attention numerics
kaix-nv
committed
Add split-K NVFP4 decode attention
kaix-nv
committed
Match NVFP4 attention fake quant to native numerics
kaix-nv
committed
Fix NVFP4 attention tests
kaix-nv
committed
Support FlashInfer vLLM attention backend
kaix-nv
committed
Document minimal MNI attention autotune design
kaix-nv
committed
Document unified vLLM attention worker design
kaix-nv
committed
Unify vLLM attention worker entry points
kaix-nv
committed
Consolidate quant and sparse attention workers
kaix-nv
committed
Honor forced ModelOpt attention kernel
kaix-nv
committed
Update unified attention worker examples
kaix-nv
committed
Simplify attention autotune schedule
kaix-nv
committed
Remove internal attention design specs
kaix-nv
committed
Simplify sparse and quant attention plan validation
kaix-nv
committed
Simplify sparse attention vLLM plugin forward paths
kaix-nv
committed
Remove redundant NVFP4 attention paths
kaix-nv
committed
Consolidate attention BMM2 quantization helpers
kaix-nv
committed
Harden vLLM attention worker per review: split mixed-batch FA decode/prefill, share sparse layer validation, disable sparse cascade, require vLLM>=0.15, reject sm<8.9
kaix-nv
committed
Refactor vLLM attention orchestration
kaix-nv
committed
Add per-operand FP8 attention quant formats for Q/K/P/V
kaix-nv
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Add nvfp4 attention support for vLLM serving#1898

Add nvfp4 attention support for vLLM serving#1898
kaix-nv merged 24 commits into
mainfrom
kaix/sparse_attn_quant_compact

Commits on Jul 15, 2026

Add compact vLLM attention quant carrier

Add compact NVFP4 V attention path

Add compact NVFP4 vLLM attention worker

Document compact NVFP4 vLLM attention serving

Harden compact NVFP4 attention numerics

Add split-K NVFP4 decode attention

Match NVFP4 attention fake quant to native numerics

Fix NVFP4 attention tests

Support FlashInfer vLLM attention backend

Document minimal MNI attention autotune design

Document unified vLLM attention worker design

Unify vLLM attention worker entry points

Consolidate quant and sparse attention workers

Honor forced ModelOpt attention kernel

Update unified attention worker examples

Simplify attention autotune schedule

Remove internal attention design specs

Simplify sparse and quant attention plan validation

Simplify sparse attention vLLM plugin forward paths

Remove redundant NVFP4 attention paths

Consolidate attention BMM2 quantization helpers

Harden vLLM attention worker per review: split mixed-batch FA decode/prefill, share sparse layer validation, disable sparse cascade, require vLLM>=0.15, reject sm<8.9

Refactor vLLM attention orchestration

Add per-operand FP8 attention quant formats for Q/K/P/V