onnxruntime
[WebGPU] Tile large block-FP8 QMoE expert matmuls
#33075
Merged
Overview
Commits
27
Changes
View On GitHub

[WebGPU] Tile large block-FP8 QMoE expert matmuls #33075

kunal-vaishnavi merged 27 commits into main from sayanshaw/qmoe-fp8-expert-matmul
sayanshaw24
Add WebGPU block-scaled FP8 QMoE
4596b55f
sayanshaw24 Correct WebGPU block-FP8 QMoE dispatch and validation
e4b8a5a3
sayanshaw24 Handle unaligned block-FP8 expert slices and test scale tiles
b4ec969e
sayanshaw24 Merge main and regenerate FP8 QMoE operator documentation
1dd47db1
sayanshaw24 Avoid repeated FP8 scale loads and cache dequant LUT
33bd8e2a
sayanshaw24 Validate canonical FP8 expert layouts and numerical decoding
f305380f
sayanshaw24 Preserve QMoE source file permissions
d436f366
kunal-vaishnavi kunal-vaishnavi requested a review from copilot-pull-request-reviewer copilot-pull-request-reviewer 20 days ago
copilot-pull-request-reviewer
copilot-pull-request-reviewer commented on 2026-10-03
sayanshaw24 Cover block-FP8 QMoE separate FC3 decode
b68ad953
sayanshaw24 Tile large block-FP8 QMoE expert matmuls on WebGPU
e44c9012
sayanshaw24 sayanshaw24 force pushed from 1413e182 to e44c9012 17 days ago
sayanshaw24 Merge remote-tracking branch 'origin/main' into sync-qmoe-fp8-32794
b3cd79c1
sayanshaw24 Merge commit 'b3cd79c10b' into sayanshaw/qmoe-fp8-expert-matmul
5c1c7ffb
sayanshaw24 Clarify QMoE packed alignment dimensions
a1c04af2
sayanshaw24 Merge commit 'a1c04af2c6' into sayanshaw/qmoe-fp8-expert-matmul
be42387e
apsonawane
apsonawane commented on 2026-10-06
kunal-vaishnavi
kunal-vaishnavi commented on 2026-10-06
kunal-vaishnavi
kunal-vaishnavi commented on 2026-10-06
kunal-vaishnavi
kunal-vaishnavi commented on 2026-10-06
kunal-vaishnavi
kunal-vaishnavi commented on 2026-10-06
Preserve FP32 block scales in tiled FP8 expert matmul
22e01960
sayanshaw24 Merge main and guard WebGPU block-FP8 projection addressing
40cebdee
Merge updated WebGPU block-FP8 QMoE base into tiled experts
a3622af1
apsonawane
apsonawane dismissed these changes on 2026-10-06
sayanshaw24 Reject orphan FP8 FC3 scales and refresh QMoE docs
09737a35
sayanshaw24 Merge reviewed FP8 QMoE FC3 validation and docs into tiled path
23922ce3
sayanshaw24 sayanshaw24 dismissed their stale review via 23922ce3 16 days ago
sayanshaw24 sayanshaw24 requested a review from apsonawane apsonawane 16 days ago
sayanshaw24 Reject unused activation scales for WebGPU block-FP8 QMoE
eb51a28f
sayanshaw24 Merge FP8 QMoE activation-scale validation into tiled path
158c7d0a
sayanshaw24 Merge current main with CUDA GQA build fix
6030cc6a
sayanshaw24 Merge current main and CUDA GQA CI fix into tiled FP8 stack
541e5c76
sayanshaw24 Merge main with ONNX materialized-node test fix
fdf7f9d5
sayanshaw24 Merge QMoE base with upstream ONNX test fix into tiled path
79873aeb
Base automatically changed from sayanshaw/qwen38-webgpu to main 15 days ago
Copilot Merge main after routed FP8 QMoE landed
41568c90
apsonawane apsonawane requested a review from copilot-pull-request-reviewer copilot-pull-request-reviewer 14 days ago
copilot-pull-request-reviewer
copilot-pull-request-reviewer commented on 2026-10-08
sayanshaw24 Merge remote-tracking branch 'origin/main' into sync/qmoe-fp8-expert-…
0cc1d4bc
sayanshaw24 Fix WebGPU FP8 QMoE subgroup tile ownership
7567ecd2
sayanshaw24
kunal-vaishnavi
kunal-vaishnavi approved these changes on 2026-10-09
kunal-vaishnavi kunal-vaishnavi merged 29feb3ba into main 13 days ago
kunal-vaishnavi kunal-vaishnavi deleted the sayanshaw/qmoe-fp8-expert-matmul branch 13 days ago

Login to write a write a comment.

Login via GitHub

Reviewers
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Terms Of Use Privacy Policy Refund Policy Impressum