[Quant] Phase 1 (video-gen): ModelOpt FP8 for HunyuanVideo-1.5 and Wan2.2#2924

Closed

lishunyang12 wants to merge 25 commits into

vllm-project:mainfrom

lishunyang12:modelopt-fp8-hv15

Commits on Apr 19, 2026

fix

roG0d
authored and
lishunyang12
committed
fix

roG0d
authored and
lishunyang12
committed
refactoring

baonudesifeizhai
authored and
lishunyang12
committed
refactoring

baonudesifeizhai
authored and
lishunyang12
committed
continue refacoring

baonudesifeizhai
authored and
lishunyang12
committed
fix huawei

baonudesifeizhai
authored and
lishunyang12
committed
fix online server problem

baonudesifeizhai
authored and
lishunyang12
committed
[Quant] Wire quant_config through HunyuanVideo-1.5 DiT (extracted from #2920)
lishunyang12
committed
[Quant] ModelOpt FP8 calibration script + stage config for HunyuanVideo-1.5
lishunyang12
committed
[Quant] HV-1.5 calibration: handle Guider config (fall back if guidance_scale kwarg unsupported)
lishunyang12
committed
[Quant] Add check_modelopt_fp8_export.py post-calibration verifier
lishunyang12
committed
[Quant] check script: read safetensors header for dtype (not get_tensor view)
lishunyang12
committed
[Quant] Force FP8 weight serialization + patch quant_algo for HV-1.5
lishunyang12
committed
[Quant] hide_quantizers_from_state_dict: pass transformer (nn.Module), not pipeline
lishunyang12
committed
[Quant] Fix calibrator's 'next' hint: text_to_video.py uses --quantization fp8, not --stage-configs-path
lishunyang12
committed
[Quant] Add --weight-block-size to switch HV-1.5 ModelOpt FP8 to per-block
lishunyang12
committed
[Quant] check script: classify weight_scale granularity (per-tensor vs per-block)
lishunyang12
committed
[Quant] check: count only meaningful (>1) dims when classifying scale granularity
lishunyang12
committed
[Quant] Wire quant_config through Wan2.2 DiT (extracted from #2920)
lishunyang12
committed
[Quant] ModelOpt FP8 calibration script + stage config for Wan2.2 TI2V-5B
lishunyang12
committed
[Quant] Fix Wan2.2 ModelOpt FP8 FFN scale-name remap (ffn.net.0 -> ffn.net_0)
lishunyang12
committed
[Quant] ModelOpt FP8 adapter: log first 3 skipped scales for diagnostics
lishunyang12
committed
[Quant] Walk submodules when collecting hf_to_vllm_mapper for ModelOpt FP8 adapter
lishunyang12
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Quant] Phase 1 (video-gen): ModelOpt FP8 for HunyuanVideo-1.5 and Wan2.2#2924

[Quant] Phase 1 (video-gen): ModelOpt FP8 for HunyuanVideo-1.5 and Wan2.2#2924
lishunyang12 wants to merge 25 commits into
vllm-project:mainfrom
lishunyang12:modelopt-fp8-hv15

Commits on Apr 19, 2026

fix

fix

refactoring

refactoring

continue refacoring

fix huawei

fix online server problem

[Quant] Wire quant_config through HunyuanVideo-1.5 DiT (extracted from #2920)

[Quant] ModelOpt FP8 calibration script + stage config for HunyuanVideo-1.5

[Quant] HV-1.5 calibration: handle Guider config (fall back if guidance_scale kwarg unsupported)

[Quant] Add check_modelopt_fp8_export.py post-calibration verifier

[Quant] check script: read safetensors header for dtype (not get_tensor view)

[Quant] Force FP8 weight serialization + patch quant_algo for HV-1.5

[Quant] hide_quantizers_from_state_dict: pass transformer (nn.Module), not pipeline

[Quant] Fix calibrator's 'next' hint: text_to_video.py uses --quantization fp8, not --stage-configs-path

[Quant] Add --weight-block-size to switch HV-1.5 ModelOpt FP8 to per-block

[Quant] check script: classify weight_scale granularity (per-tensor vs per-block)

[Quant] check: count only meaningful (>1) dims when classifying scale granularity

[Quant] Wire quant_config through Wan2.2 DiT (extracted from #2920)

[Quant] ModelOpt FP8 calibration script + stage config for Wan2.2 TI2V-5B

[Quant] Fix Wan2.2 ModelOpt FP8 FFN scale-name remap (ffn.net.0 -> ffn.net_0)

[Quant] ModelOpt FP8 adapter: log first 3 skipped scales for diagnostics

[Quant] Walk submodules when collecting hf_to_vllm_mapper for ModelOpt FP8 adapter

Commits on May 17, 2026

Phase1 (video-gen) ModelOpt FP8 Follow-ups (#57)

Merge upstream/main into modelopt-fp8-hv15