What does this PR do?

Type of change: New example feature

Add Wan 2.2 T2V A14B NVFP4-SVDQuant HuggingFace checkpoint export to the diffusers quantization example, mirroring the Qwen-Image SVDQuant mechanics (#1706):

• --model wan2.2-t2v-14b now quantizes both experts by default — transformer (high-noise) first, then transformer_2 (low-noise), so the low-noise expert calibrates against the already-quantized high-noise expert, matching deployment.
• The default Wan recipe (filter_func_wan_video: exclude the first 3 / last 3 of the 40 blocks, plus everything outside blocks) is now applied before calibration via the block_range mechanism. This is required for SVDQuant: the algorithm folds AWQ smoothing scales and subtracts the low-rank residual from every enabled linear during calibration, so post-hoc disabling (the previous filter-func-only path) would leave "excluded" weights silently mutated. With this PR they stay bit-identical to the source.
• VAE backbones (--backbone vae) skip the transformer block-range recipe and keep their dedicated Conv3D recipe.
• No svdquant_skip_layers for Wan: every quantized linear (self-attn, cross-attn incl. to_k/to_v, FFN) carries the full SVDQuant recipe (pre_quant_scale + rank-r svdquant_lora_a/b).
• ONNX export paths untouched.

Usage

python examples/diffusers/quantization/quantize.py \
    --model wan2.2-t2v-14b \
    --model-dtype BFloat16 --trt-high-precision-dtype BFloat16 \
    --format fp4 --quant-algo svdquant --lowrank 32 \
    --batch-size 1 --calib-size 16 --n-steps 20 \
    --hf-ckpt-dir ./hf_ckpt

Testing

• New wan22_14b_nvfp4_svdquant case in tests/examples/diffusers/test_export_diffusers_hf_ckpt.py asserting per expert (transformer and transformer_2): NVFP4_SVD quant config + lora_rank; only blocks {3, 4} of the 8-block tiny fixture quantized and nothing outside blocks; svdquant_lora_a/b + pre_quant_scale present on all 10 block linears with rank-consistent shapes; NVFP4 weight_scale_2; no live-quantizer keys leaked.
• Tiny Wan fixture updated: num_layers 2→8 (the block-range recipe needs ≥ 8 blocks), heads 2→4 (hidden 48, divisible by the NVFP4 block size 16; head_dim stays 12 so the RoPE axis split stays even).
• All 11 affected example tests pass on GB200 (sm100), 653 s total: 3 Wan HF-export cases (int8 / fp8 / new svdquant), 7 Wan quantize+restore cases incl. --backbone vae fp8/fp4 (verifying the VAE gate), and the qwen_nvfp4_svdquant regression.
• Real-model smoke: Wan2.2-T2V-A14B rank-32/64/128 SVDQuant exports with a structural verify gate (both experts NVFP4_SVD; blocks 3–36 quantized; excluded blocks and non-block modules bit-identical to the source BF16; full pipeline dirs). Rank-32 in progress; results will be posted on this PR.

Before your PR is "Ready for review"

• Is this change backward compatible?: ❌ — --model wan2.2-t2v-14b previously quantized only transformer by default; it now quantizes both experts (a single-expert checkpoint of the two-expert pipeline is half-quantized). Pass --backbone transformer to restore the old behavior. The tiny Wan test fixture also grew from 2 to 8 blocks / hidden 24 to 48.
• If you copied code from any other sources or added a new PIP dependency, did you follow guidance in CONTRIBUTING.md: N/A
• Did you write any new necessary tests?: ✅
• Did you update Changelog?: ✅
• Did you get Claude approval on this PR?: ❌ (draft)

Additional Information

Draft until the real-model rank-32/64/128 smoke exports and verify gates complete.

🤖 Generated with Claude Code

Update: AWQ-Lite search removed from the SVDQuant path

SVDQuant calibration is now SmoothQuant-style fixed migration only (SVDQuantConfig.alpha, default 1.0 = migrate outliers fully to the weights; the diffusers example reuses --alpha). The AWQ-Lite search is gone from svdquant: two forward-loop passes instead of three, no per-alpha quantized GEMM sweep. (The transient --svdquant-alpha opt-in from the second commit was superseded in the fourth commit.)

Why: the search optimized pre-SVD, weight-only output MSE with activations unquantized — a mismatched objective twice over for W4A4 SVDQuant, where the SVD low-rank branch absorbs the migrated outliers and activation flatness is the point of the migration (SVDQuant paper design).

Validation on Wan2.2-T2V-A14B (rank 32, calib 1 x 20 steps @ 720x1280x81f, GB200):

Gate checks per expert: NVFP4_SVD + lora_rank=32; only blocks 3-36 quantized; 340/340 linears with svdquant_lora_a/b + pre_quant_scale; excluded region bit-identical in export precision (the Wan-AI repo ships FP32 shards, so excluded tensors are pure BF16 downcasts - 0 value changes). Unit: 16/16 calib/mode tests incl. new test_svdquant_smoothquant_calibration (asserts 2 forward loops + the SmoothQuant scale formula). Example: Wan + Qwen SVDQuant export tests pass on the new path.

🤖 Generated with Claude Code