[Kernel][WIP] Integrate native FlashInfer SM70 batch GDN and QSA probes - #523

Draft

yangzhuxinyzx wants to merge 31 commits into

mainfrom

codex/v100-flashinfer-batch-integration-20260906-020903

Commits on Sep 5, 2026

[Kernel][WIP] Fuse batched HC pointwise and disjoint push gather

yangzhuxinyzx
and
codex
committed
0936cd2
[Benchmark] Isolate HC channels and validate canonical QSA batch routes

yangzhuxinyzx
and
codex
committed
e28351c
[Benchmark] Screen HC weight views and reject slow QSA batch routes

yangzhuxinyzx
and
codex
committed
9c0a859
[Kernel] Integrate opt-in batched HC with isolated TP channel

yangzhuxinyzx
and
codex
committed
aeba8e2
[Benchmark] Add concurrent tool and schema quality coverage

yangzhuxinyzx
and
codex
committed
b64dae6
[Benchmark] Record full-model HC gains and remaining quality gates

yangzhuxinyzx
and
codex
committed
30837ca
[Benchmark] Reject malformed tool termination and record spawn-safe retry

yangzhuxinyzx
and
codex
committed
0ed451f
[Benchmark] Correct BFCL dictionary scoring and reject prefill-only trace

yangzhuxinyzx
and
codex
committed
b1d3bd1
[Benchmark] Attribute full C16 graphs and screen QSA grid amortization

yangzhuxinyzx
and
codex
committed
3f1fa2d
[Benchmark] Record QSA long-context regressions and resource limits

yangzhuxinyzx
and
codex
committed
a5bceb3
[Kernel] Fuse batched GDN projection copies without changing GEMMs

yangzhuxinyzx
and
codex
committed
eb40e64
[Benchmark] Compare original batch tool quality and retain adverse signal

yangzhuxinyzx
and
codex
committed
b6c18db
[Benchmark] Retain batch quality ablations and reject single-tile MoE

yangzhuxinyzx
and
codex
committed
46f2b79
[Benchmark] Validate QSA order repair and isolate sparse batch overhead

yangzhuxinyzx
and
codex
committed
5a04923
[Kernel] Prototype FlashInfer CUDA sparse QSA decode on SM70

yangzhuxinyzx
and
codex
committed
6b2f4ad
[Perf] Record FlashInfer SM70 QSA graph and sanitizer screening

yangzhuxinyzx
and
codex
committed
17fa36e
[Kernel][WIP] Adapt FlashInfer GDN and HC layer fusion to SM70

yangzhuxinyzx
and
codex
committed
e2061d5
[Test] Screen independent FlashInfer GDN histories and HC semantics

yangzhuxinyzx
and
codex
committed
72af224

Commits on Sep 6, 2026

[Kernel][WIP] Screen fused GDN gains and register-staged HC

yangzhuxinyzx
and
codex
committed
9ea54c1
[Integration][WIP] Combine isolated FlashInfer and batch HC candidates

yangzhuxinyzx
and
codex
committed
f8e822c
[Kernel][WIP] Bridge FlashInfer batch GDN and QSA into model execution

yangzhuxinyzx
and
codex
committed
190e5f0
[Bugfix][WIP] Keep FlashInfer GDN dispatch inside the compiled boundary
yangzhuxinyzx
committed
45711b5
[Bugfix][WIP] Resolve AOT state placeholders before fused GDN capture
yangzhuxinyzx
committed
2273c24
[Test] Preserve dynamic environment lookup in FlashInfer bridge fixtures
yangzhuxinyzx
committed
f45c673
[Doc] Record combined FlashInfer throughput and failed tool-quality gate
yangzhuxinyzx
committed
25fd594
[Doc] Record FlashInfer quality attribution and control variability
yangzhuxinyzx
committed
a25c23f
[Kernel][WIP] Add device-planned SM70 MQA scoring and quality localization

yangzhuxinyzx
and
codex
committed
64dd679
[Bugfix][SM70] Stabilize QSA page4 order under KV relocation

Leonccaa
authored and
yangzhuxinyzx
committed
3e0f7a4
[Kernel][WIP] Stabilize QSA planning and package SM70 GDN

yangzhuxinyzx
and
codex
committed
d1c7e64
[Test][WIP] Isolate sparse QSA rounding and HC projection stages

yangzhuxinyzx
and
codex
committed
1a5a7f7
[Bugfix] Preserve HC down arithmetic and prototype compatible SM70 sparse QSA

yangzhuxinyzx
and
codex
committed
6dbb629

Navigation Menu

Uh oh!

[Kernel][WIP] Integrate native FlashInfer SM70 batch GDN and QSA probes - #523

[Kernel][WIP] Integrate native FlashInfer SM70 batch GDN and QSA probes#523
yangzhuxinyzx wants to merge 31 commits into
mainfrom
codex/v100-flashinfer-batch-integration-20260906-020903

Commits on Sep 5, 2026

[Kernel][WIP] Fuse batched HC pointwise and disjoint push gather

[Benchmark] Isolate HC channels and validate canonical QSA batch routes

[Benchmark] Screen HC weight views and reject slow QSA batch routes

[Kernel] Integrate opt-in batched HC with isolated TP channel

[Benchmark] Add concurrent tool and schema quality coverage

[Benchmark] Record full-model HC gains and remaining quality gates

[Benchmark] Reject malformed tool termination and record spawn-safe retry

[Benchmark] Correct BFCL dictionary scoring and reject prefill-only trace

[Benchmark] Attribute full C16 graphs and screen QSA grid amortization

[Benchmark] Record QSA long-context regressions and resource limits

[Kernel] Fuse batched GDN projection copies without changing GEMMs

[Benchmark] Compare original batch tool quality and retain adverse signal

[Benchmark] Retain batch quality ablations and reject single-tile MoE

[Benchmark] Validate QSA order repair and isolate sparse batch overhead

[Kernel] Prototype FlashInfer CUDA sparse QSA decode on SM70

[Perf] Record FlashInfer SM70 QSA graph and sanitizer screening

[Kernel][WIP] Adapt FlashInfer GDN and HC layer fusion to SM70

[Test] Screen independent FlashInfer GDN histories and HC semantics

Commits on Sep 6, 2026

[Kernel][WIP] Screen fused GDN gains and register-staged HC

[Integration][WIP] Combine isolated FlashInfer and batch HC candidates

[Kernel][WIP] Bridge FlashInfer batch GDN and QSA into model execution

[Bugfix][WIP] Keep FlashInfer GDN dispatch inside the compiled boundary

[Bugfix][WIP] Resolve AOT state placeholders before fused GDN capture

[Test] Preserve dynamic environment lookup in FlashInfer bridge fixtures

[Doc] Record combined FlashInfer throughput and failed tool-quality gate

[Doc] Record FlashInfer quality attribution and control variability

[Kernel][WIP] Add device-planned SM70 MQA scoring and quality localization

[Bugfix][SM70] Stabilize QSA page4 order under KV relocation

[Kernel][WIP] Stabilize QSA planning and package SM70 GDN

[Test][WIP] Isolate sparse QSA rounding and HC projection stages

[Bugfix] Preserve HC down arithmetic and prototype compatible SM70 sparse QSA