[inureyes]

Closes #2444

nvidia/Qwen3-30B-A3B-NVFP4 (ModelOpt) failed to load on main with 1-bit weights are affine-only ... (prefix: model.layers.0.self_attn.q_proj). The fused QKV loader hardcoded affine, so the bias-free nvfp4 planes solved to a 1-bit affine layout, and the concatenated plane had no slot for each plane's weight_scale_2.

Changes

• FusedQKVLinear::from_weights_separate infers the mode from q_proj with infer_quantization_mode, as UnifiedLinear::from_weights does.
  • If any of q/k/v has a {plane}.global_scale sidecar, the layer loads as QkvProjection::Split. Each plane is built by UnifiedLinear::from_weights_with_mode at its reconciled layout, applies its scalar after the matmul (key name and meaning unchanged) and claims its own width-pinned runtime-LoRA terms (fix(lora): pin runtime LoRA term widths on quantized linear bases #2438).
  • The fused arm reconciles a block-float plane's group size from its shapes and checks the backend before storing it. Previously it stored the declared value, for example 64 for an mxfp4 plane packed at 32.
  • No per-layer struct gains a field (feat(lora): support mlx-lm LoRAEmbedding adapters on embed_tokens #2434).
• Qwen3 and Qwen3-MoE group_size() fall back to modelopt_nvfp4_group_size (16), as PhiMoE does. This adds a #[serde(default)] quantization_config field to both ModelArgs and one line to each of three struct-literal test fixtures.
• finish_vlm_weights_common and load_gemma4_unified (raw keys, before the sanitizer) run the native global-scale flag pass.
  • The filtered VLM loaders end in finish_vlm_weights_common too. There a subset variant skips a flagged plane with neither its weight nor its scale in the map. A loaded flagged plane without its scale still fails.
  • split_fused_moe_key handles .native_global_scale and .global_scale. The [num_experts] vector is copied unsplit to gate_proj and up_proj and renamed under down_proj, so an unflagged scale fails at its switch_glu.gate_proj prefix instead of being orphaned.
• DeepSeek v1 SwitchLinear::from_stacked_parts refuses a plane that has only one of .scales / .biases, naming the prefix. Previously it built Regular and aborted in gather_mm.

Validation (GB10, CUDA, every GPU run under gpu-lock)

• MLXCEL_SDPA_DETERMINISTIC=1 mlxcel generate -m ~/models_archive/hf-src/Qwen3-30B-A3B-NVFP4 -p "The capital of France is" -n 80 --temp 0 --show-reasoning loads (main fails at load) and generates coherent text: "... Since the question is about the capital of France, the answer is Paris. ...". No mlx-lm teacher-forced trace: mlx-lm 0.32.2 refuses this checkpoint (55968 unexpected parameters, unstacked ModelOpt keys).
• Greedy output is token-identical to main (MLXCEL_SDPA_DETERMINISTIC=1, same prompts) for ~/models_archive/mlx/laguna-xs-2.1-nvfp4 (with --show-reasoning), olmoe-1b-7b-0924-nvfp4-gs and gemma-4-31b-it-nvfp4 (two prompts). This was re-checked on the final commit.
• 10 new tests:
  • layers: nvfp4 split QKV with per-plane scales, its forward checked against x @ dequantize(w).T * s; mxfp4 QKV fused at group 32.
  • qwen3 / qwen3_moe: ModelOpt-only group_size() == 16.
  • Gemma 4 Unified: flagged and unflagged fused experts.
  • VLM: a shipped .native_global_scale is refused, a flagged scale is resolved, and a filtered load skips planes outside the filter.
  • DeepSeek: the half-quantized plane.
  • Each test fails with its fix reverted and passes with it restored.
• These filters pass: layers::, runtime_lora, one_bit (mlxcel-core) and qwen3 vlm deepseek:: switch_layers sanitize (mlxcel). Clippy --lib --tests -D warnings is clean for both crates. fmt and the five contract checks pass.

Not addressed here (pre-existing): the Kimi K3 VLM loader bypasses finish_vlm_weights_common, so a flagged export is refused cleanly at SwitchLinear. A dense .global_scale is not shape-validated, which is the same as o_proj and MLP on main.