ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
unsloth/scripts
History
Daniel Han 42a4c07c3c Make the video speed/mem bench MoE-aware; add A14B + Hunyuan-720p families 
Adds wan2.2-t2v-a14b and hunyuanvideo-1.5-720p to the bench family table and makes _apply_levers
quantize / compile BOTH experts of a dual-expert MoE (Wan2.2-A14B) via a _SecondExpertView proxy
that mirrors the loader's _SecondDiTView, so A14B latency and accuracy are measured on the real
two-DiT path instead of only the first expert.

Used to validate that every video family is on the fastest DiT quant scheme that is not less
accurate than its alternative (B200, 512x320, 25 frames, 30 steps, no cache, LPIPS vs dense bf16):
- Wan2.2-TI2V-5B  fp8 49.9 ms/step vs int8 64.6 vs dense 59.8; LPIPS 0.129 vs 0.180
- Wan2.2-T2V-A14B fp8 195.8 ms/step vs int8 193.5 vs dense 369.1; LPIPS 0.288 vs 0.349 (both experts)
- LTX-2 / LTX-2.3  fp8 133.5 ms/step vs int8 138.4 vs dense 204.7; LPIPS 0.026 vs 0.027
- HunyuanVideo-1.5 fp8 is black (denied, not localizable); int8 21.2 s vs dense+compile 15.4 s -- a
  memory saving at a speed cost, so int8 stays only because fp8 is impossible there.

fp8 is faster than dense AND more accurate than int8 on all three Wan/LTX families (the two Wan ones
via the condition_embedder exclude; LTX-2 needs none -- its conditioning has no zero-amax padding
rows). The auto-ladder + deny + exclude already select exactly these schemes, so no scheme-selection
change was needed; this commit is the bench faithfulness improvement that let the campaign confirm it.
2026-07-09 01:33:56 +00:00
..
data CI: scope GITHUB_TOKEN permissions, add MLX CI, unblock ~60 skipped tests (#5312) 2026-05-11 03:19:13 -07:00
build_prequant_checkpoint.py Restore fp8 DiT quant for Wan video via a per-family embedder exclude 2026-07-08 23:22:34 +00:00
check_frontend_dep_removal.py Reduce and tighten code comments and docstrings repo-wide (#6095) 2026-06-08 23:09:51 -07:00
check_new_install_scripts.py Reduce and tighten code comments and docstrings repo-wide (#6095) 2026-06-08 23:09:51 -07:00
compare_engines.py Studio diffusion (Phase 7): accuracy-preserving speed pass (2.2x via GGUF compile) (#6690) 2026-07-01 15:31:50 -03:00
compile_probe.py Studio diffusion (Phase 7): accuracy-preserving speed pass (2.2x via GGUF compile) (#6690) 2026-07-01 15:31:50 -03:00
diffusion_bench.py Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source) (#6694) 2026-07-01 15:36:43 -03:00
diffusion_quality.py [pre-commit.ci] auto fixes from pre-commit.com hooks 2026-07-02 03:30:16 +00:00
enforce_kwargs_spacing.py Reduce and tighten code comments and docstrings repo-wide (#6095) 2026-06-08 23:09:51 -07:00
fbcache_flux_probe.py Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT (#6703) 2026-07-01 15:40:52 -03:00
fp8_layer_ablation.py Restore fp8 DiT quant for Wan video via a per-family embedder exclude 2026-07-08 23:22:34 +00:00
fp8_overflow_check.py Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source) (#6694) 2026-07-01 15:36:43 -03:00
install_gemma4_mlx.sh Update Install Scripts (#5968) 2026-06-03 05:39:42 -07:00
install_qwen3_6_mlx.sh Update Install Scripts (#5968) 2026-06-03 05:39:42 -07:00
install_rocm_wsl_strixhalo.sh ROCm-on-WSL: support discrete Radeon (RDNA 3/4) in WSL, not just Strix Halo (#6915) 2026-07-07 02:29:37 -07:00
int8_linear_probe.py Studio diffusion (Phase 14): fix int8 dense quant on Flux / Qwen (skip M=1 modulation linears) (#6716) 2026-07-01 15:41:49 -03:00
leverage_probe.py Studio diffusion (Phase 7): accuracy-preserving speed pass (2.2x via GGUF compile) (#6690) 2026-07-01 15:31:50 -03:00
lint_workflow_triggers.py Reduce and tighten code comments and docstrings repo-wide (#6095) 2026-06-08 23:09:51 -07:00
lockfile_supply_chain_audit.py Reduce and tighten code comments and docstrings repo-wide (#6095) 2026-06-08 23:09:51 -07:00
notebook_to_python.py Reduce and tighten code comments and docstrings repo-wide (#6095) 2026-06-08 23:09:51 -07:00
notebook_validator.py Reduce and tighten code comments and docstrings repo-wide (#6095) 2026-06-08 23:09:51 -07:00
nvfp4_probe.py Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source) (#6694) 2026-07-01 15:36:43 -03:00
nvfp4_t211_probe.py Fix diffusion training validation, dataset upload atomicity, and LoRA error mapping 2026-07-04 03:17:12 -03:00
perf_levers_probe.py [pre-commit.ci] auto fixes from pre-commit.com hooks 2026-07-02 03:30:16 +00:00
perf_verify.py Studio diffusion (Phase 7): accuracy-preserving speed pass (2.2x via GGUF compile) (#6690) 2026-07-01 15:31:50 -03:00
prequant_probe.py Studio diffusion (Phase 9): pre-quantized transformer loading (#6700) 2026-07-01 15:37:53 -03:00
quant_accuracy_sweep.py Gate VAE fp8 quant by decoded-image accuracy (auto layerwise fp8, fp8_dynamic opt-in) 2026-07-08 10:45:27 +00:00
quant_probe.py Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source) (#6694) 2026-07-01 15:36:43 -03:00
quant_speedmem_bench.py Add video speed/memory lever benchmark; extend quant bench with DiT mode 2026-07-08 14:35:58 +00:00
run_ruff_format.py Reduce and tighten code comments and docstrings repo-wide (#6095) 2026-06-08 23:09:51 -07:00
scan_npm_packages.py scan_packages: key baseline on matched-code hash so payloads in baselined files are not auto-suppressed (#6552) 2026-07-01 04:03:59 -07:00
scan_npm_packages_baseline.json scan_packages: key baseline on matched-code hash so payloads in baselined files are not auto-suppressed (#6552) 2026-07-01 04:03:59 -07:00
scan_packages.py CI: baseline the fastapi/gguf scan findings for the studio dependency set 2026-07-02 02:04:28 +00:00
scan_packages_baseline.json Merge remote-tracking branch 'origin/main' into fold-integration 2026-07-07 05:52:16 +00:00
sd_cpp_smoke.py Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine (#6680) 2026-07-01 15:18:38 -03:00
sparse_accum_probe.py Fix diffusion training validation, dataset upload atomicity, and LoRA error mapping 2026-07-04 03:17:12 -03:00
stamp_studio_release.py Reduce and tighten code comments and docstrings repo-wide (#6095) 2026-06-08 23:09:51 -07:00
sync_allow_scripts_pins.py Studio: auto-sync allowScripts pins after dependency bumps (#6136) 2026-06-10 02:35:37 -07:00
uninstall.ps1 Merge image-generation bug fixes (#6872) 2026-07-07 16:14:06 +00:00
uninstall.sh Keep an unowned default-mode sd.cpp checkout on uninstall 2026-07-07 10:22:40 +00:00
verify_comment_only_diff.py Reduce and tighten code comments and docstrings repo-wide (#6095) 2026-06-08 23:09:51 -07:00
verify_import_hoist.py studio: tool calling for DeepSeek (R1/V3/V3.1), GLM 4.x, Kimi K2 on safetensors + MLX (#5624) 2026-07-06 15:40:46 -07:00
verify_prequant_backend.py [pre-commit.ci] auto fixes from pre-commit.com hooks 2026-07-02 03:30:16 +00:00
video_quality.py Fail video quality gate on candidate frame-count mismatch 2026-07-05 07:46:42 +00:00
video_speedmem_bench.py Make the video speed/mem bench MoE-aware; add A14B + Hunyuan-720p families 2026-07-09 01:33:56 +00:00