[OMNIML-6073] [2/3] Quantize fused experts under grouped_mm; leave transposed layouts unquantized by shengliangxu · Pull Request #2745 · NVIDIA/Model-Optimizer

Quantize fused experts under grouped_mm by intercepting _grouped_linear

shengliangxu

committed

5a329b8

Leave fused experts stored transposed unquantized

shengliangxu

committed

54b50f5

Keep a disabled weight quantizer's pre-quant scale and rotation on the grouped forward

shengliangxu

committed

2d287cf

Pin the grouped forward's fake-quantized output and gradients against eager

shengliangxu

committed

144db5a

Unregister the grouped test experts in a finally, as the class's other tests do

shengliangxu

committed

9122ae1

Leave the torch_a2a test fixtures to the PR whose tests drive them

shengliangxu

committed

6372ed9

Fake-quantize the grouped forward's experts in one launch where the format allows

shengliangxu

committed

7c83909

Let the eager-keyed weight-input captures skip the grouped forward's batch

shengliangxu

committed

8c84ef9

Take the stacked fake-quant kernel on the grouped forward without autograd too

shengliangxu

committed

c7fb068

Name expert parallelism in the eager-only wrapper's grouped-forward refusal

shengliangxu

committed

123e5f7

Say the grouped forward's weight-input captures skip its batch

shengliangxu

committed

fb41c9e

Pin that a disabled weight quantizer's rotation applies on the grouped forward

shengliangxu

committed

4b5bc94

Drop a comment that repeats the docstring, and a move of a helper

shengliangxu

committed

791ce88

Say in the changelog that an old modelopt_state for such experts no longer restores

shengliangxu

committed

b68e15d

Import transformers' moe module in the GPU test as the plugin does

shengliangxu

committed

d0c695d

Fake-quantize the grouped weight's slices in strict step with its quantizers

shengliangxu

committed

814f60c

Keep the quantizers' dtype in the grouped forward without autograd

shengliangxu

committed

57e12d7

Pin the export refusal and the restore break for transposed experts

shengliangxu

committed

4f1b64a

File the restore break under Backward Breaking Changes

shengliangxu

committed

f85148d

Promote the grouped weight's buffer as the stack does when experts disagree

shengliangxu

committed

77c6f7c

Say the grouped_mm fallback does not call F.linear

shengliangxu

committed

30d5750

Say why calibration is still forced onto eager now that grouped_mm is intercepted

shengliangxu

committed

5f9739a

Say what keeps the FP8 experts wrapper on eager

shengliangxu

committed

689d372

Say the transposed check sees only a declared non-gated layout

shengliangxu

committed

1274887

Count the FSDP2 ownership scans themselves, not fully_shard.state calls

shengliangxu

committed

3cfe6e0

Say why transposed experts stay in the eager scan and that their export is refused

shengliangxu

committed

75265cb

Navigation Menu

Repository navigation

[OMNIML-6073] [2/3] Quantize fused experts under grouped_mm; leave transposed layouts unquantized - #2745

[OMNIML-6073] [2/3] Quantize fused experts under grouped_mm; leave transposed layouts unquantized#2745
shengliangxu merged 26 commits into
mainfrom
shengliangx/hfep-8-fused-experts

Commits on Oct 28, 2026

Quantize fused experts under grouped_mm by intercepting _grouped_linear

Leave fused experts stored transposed unquantized

Keep a disabled weight quantizer's pre-quant scale and rotation on the grouped forward

Pin the grouped forward's fake-quantized output and gradients against eager

Unregister the grouped test experts in a finally, as the class's other tests do

Leave the torch_a2a test fixtures to the PR whose tests drive them

Fake-quantize the grouped forward's experts in one launch where the format allows

Let the eager-keyed weight-input captures skip the grouped forward's batch

Take the stacked fake-quant kernel on the grouped forward without autograd too

Name expert parallelism in the eager-only wrapper's grouped-forward refusal

Say the grouped forward's weight-input captures skip its batch

Pin that a disabled weight quantizer's rotation applies on the grouped forward

Drop a comment that repeats the docstring, and a move of a helper

Say in the changelog that an old modelopt_state for such experts no longer restores

Import transformers' moe module in the GPU test as the plugin does

Fake-quantize the grouped weight's slices in strict step with its quantizers

Keep the quantizers' dtype in the grouped forward without autograd

Pin the export refusal and the restore break for transposed experts

File the restore break under Backward Breaking Changes

Promote the grouped weight's buffer as the stack does when experts disagree

Say the grouped_mm fallback does not call F.linear

Say why calibration is still forced onto eager now that grouped_mm is intercepted

Say what keeps the FP8 experts wrapper on eager

Say the transposed check sees only a declared non-gated layout

Count the FSDP2 ownership scans themselves, not fully_shard.state calls

Say why transposed experts stay in the eager scan and that their export is refused