pipelines: forward --quantization to the multi-GPU runners

[Boning (wwelder)]

Problem

The three multi-GPU CLIs — distilled_mgpu.py, ti2vid_two_stages_mgpu.py, ti2vid_two_stages_hq_mgpu.py — never pass quantization to controller.start(). setup() therefore always takes its fallback:

quantization_policy = (
    quantization() if quantization is not None else _build_fp8_cast_policy(model_paths.transformer())
)

so --quantization is silently ignored on every multi-GPU path. The single-GPU pipelines do forward it (quantization=args.quantization in distilled.py), so the flag works there.

With an NVFP4 checkpoint the run then fails inside the fp8 loader, which expects a scalar scale and rejects NVFP4's per-block scales:

SymmetricRunnerError: Unsupported scale shape (4096, 256) for transformer_blocks.0.attn1.to_k.weight

Reproduce with the documented nvfp4-prequant invocation, but on the mgpu entry point:

python -m ltx_pipelines.distilled_mgpu \
    --transformer-path .../ltx-2.5-22b-distilled-transformer-nvfp4.safetensors \
    --text-encoder-path .../gemma4-12b-with-proj-ltx-2.5-bf16.safetensors \
    --video-vae-path .../ltx-2.5-video-vae-conv-bf16.safetensors \
    --audio-vae-path .../ltx-2.5-audio-vae-bf16.safetensors \
    --spatial-upsampler-path .../ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors \
    --quantization nvfp4-prequant --prompt "..." --output-path out.mp4

Why args.quantization can't just be forwarded

setup() documents its parameter as "a picklable zero-arg builder (built per worker, post-spawn)" and types it Callable[[], QuantizationPolicy] | None, whereas args.quantization is an already-built QuantizationPolicy. The policy is constructed in the parent and would have to survive spawn; the builder deliberately runs inside each worker instead.

_resolve_quantization already has both kind and the resolved checkpoint path, but discards them after building the policy. This adds namespace.quantization_builder, a functools.partial over QuantizationKind.to_policy. Enum members and bound methods pickle by reference, so the partial round-trips through spawn cleanly.

Behaviour

• Single-GPU pipelines: unchanged, they keep reading args.quantization.
• No --quantization: quantization_builder is None, and the runners keep their existing fp8-cast default.
• _resolve_quantization runs unconditionally in the parser's parse_args override, alongside _resolve_model_paths, so the attribute always exists wherever args.model_paths already does.

Verification

Tested on 2× and 4× RTX PRO 6000 Blackwell (sm_120) with ltx-2.5-22b-distilled-transformer-nvfp4.safetensors plus the bf16 text encoder and conv video VAE. Before: every run failed with the scale-shape error above. After: the NVFP4 checkpoint loads and generates normally — dozens of 15 s clips at 576×832 and 704×1280, across sequence-parallel world sizes 2 and 4.