[WebGPU] Fuse rank-32 quantized LoRA with MatMulNBits #33370
qjia7
merged
49 commits into
microsoft:main
from
qjia7:feat/webgpu-matmul-nbits-qlora
Fix 1DS telemetry deadlock during Windows process termination
5578aabb
Add Windows child-process coverage for 1DS teardown
625b47c4
[WebGPU] Fuse rank-32 FP16 LoRA with MatMulNBits
897462aa
[WebGPU] Fuse rank-32 INT8 adapters with MatMulNBitsQLoRA
5a5d3258
Optimize WebGPU QLoRA prefill with shared DP4A input and narrow tiles
2638ae21
Batch WebGPU input uploads to amortize adapter transfer overhead
8eab6582
Optimize narrow WebGPU QLoRA projections with small-M DP4A
210584c7
Optimize WebGPU QLoRA decode with register reuse and subgroup reduction
b3e1633c
Document WebGPU LoRA performance and numerical limits
6a1cf7d3
Merge main and fix WebGPU LoRA CI compatibility
7e707993
Merge remote-tracking branch 'origin/main' into feat/webgpu-matmul-nb…
5b64b23f
Validate compact QLoRA capture and scope FP16 test tolerances
788de1b1
Fix GCC range-loop warning in QLoRA decode test
8df78337
[WebGPU] Support FP16 and FP32 models in LoRA and QLoRA
c9d36921
[WebGPU] Fuse FP32 QLoRA decode base and down projection with DP4A
4e8a5373
[WebGPU] Parallelize floating LoRA projections for prefill and decode
0243e402
WebGPU: separate floating LoRA from QLoRA support
139ad24c
Fix QLoRA fusion test template parsing on GCC 11
2ea9a834
qjia7
changed the title
[WebGPU] Generalize K-parallel DP4A dispatch for narrow outputs
9ea931eb
Merge branch 'main' into feat/webgpu-matmul-nbits-qlora
0cf697a5
[WebGPU] Enable QLoRA fast paths by device capabilities
2ea6f9bb
[WebGPU] Simplify QLoRA dispatch and reuse DP4A for short FP32 prefill
41037661
[WebGPU] Fix QLoRA CI tolerances and execute FP16 INT2 tests
51a970c2
[WebGPU] Note follow-up for arbitrary QLoRA ranks
26110961
Merge remote-tracking branch 'origin/main' into feat/webgpu-matmul-nb…
47214c46
Remove duplicate telemetry test registration after merging main
7b988026
copilot-pull-request-reviewer
commented
on 2026-10-19
[WebGPU] Adapt QLoRA decode to subgroup ID opt-in
a7900dc0
[WebGPU] Guard QLoRA down projection against padded dispatches
be2c49cf
copilot-pull-request-reviewer
commented
on 2026-10-19
webgpu: Fuse QLoRA subgroup prefill and share narrow down input
b684a410
webgpu: Batch unmapped input uploads on UMA devices
034acaaf
webgpu: Batch plugin input uploads on the owning session stream
a49f0196
webgpu: Merge updated QLoRA PR and preserve local optimizations
6cce74f5
webgpu: Merge latest main and preserve QLoRA matmul support
2971861c
webgpu: Generalize QLoRA schema and simplify correctness tests
4625103a
webgpu: Extract QLoRA shader helpers into WGSL templates
f248e412
webgpu: Disambiguate plugin upload test provider options
40fc0c4f
Support empty MatMulNBits weights on CPU and WebGPU
171cbe3e
Merge main after upstreaming WebGPU input upload batching
507021fd
Fix dependent template call in MatMulNBits empty-weight path
12d74efa
Enable WebGPU QLoRA fusion automatically and preserve defaults
5f327263
Merge main into WebGPU MatMulNBits QLoRA
93569ab3
[WebGPU] Simplify DP4A tiling for MatMulNBits QLoRA
1360fa84
Merge main into feat/webgpu-matmul-nbits-qlora
185cfd7b
[WebGPU] Refresh MatMulNBitsQLoRA documentation
38638f61
mustjab
dismissed these changes
on 2026-10-30
[WebGPU] Clean up QLoRA shader access and empty-weight handling
f261ba56
webgpu: Preserve QLoRA projection semantics
9af5cd83
webgpu: Reset QLoRA accumulator for each column group
4fe0c2ab
webgpu: Clear padding when reloading QLoRA staging tiles
d9379f78
xiaofeihan1
commented
on 2026-11-02
webgpu: Size QLoRA staging for four-wave BK=32 tiles
512827f8
mustjab
approved these changes
on 2026-11-02
xiaofeihan1
approved these changes
on 2026-11-03
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub