fix(heretic2-nvfp4): quant as ConditionalGeneration (namespace fix) + modelopt recipe for working MTP

Root-caused the NVFP4 gibberish to a quant-namespace bug: quant_nvfp4.py loaded
via AutoModelForCausalLM -> text-only Qwen3_5ForCausalLM -> flat model.layers.* keys,
but vLLM 0.24 serves only Qwen3_5ForConditionalGeneration (whose weight mapper needs
model.language_model.*). Fixed by loading as AutoModelForImageTextToText; NVFP4 now
serves coherent (validated greedy on ana-ml2 GPU0).

Base NVFP4 (compressed-tensors) measured ~53 tok/s (~= GGUF at batch-1, no single-stream
win) and its MTP is 0% acceptance (vLLM's Qwen3_5MTP drafter loads the bf16 mtp head only
off a modelopt main-model checkpoint). Added quant_modelopt.py (nvidia-modelopt PTQ,
matches AEON's NVFP4 W4A4 g16 + lm_head/linear_attn/visual exclusions) as the path to
working native MTP; graft + splice + serve otherwise unchanged.

services/heretic2-nvfp4-quant/quant_nvfp4.py

@@ -82,12 +82,24 @@ def main() -> int:
     ap.add_argument("--seqlen", type=int, default=8192)
     args = ap.parse_args()
 
-    from transformers import AutoModelForCausalLM, AutoTokenizer
+    from transformers import AutoModelForImageTextToText, AutoTokenizer
     from llmcompressor import oneshot
     from llmcompressor.modifiers.quantization import QuantizationModifier
 
     print(f"loading grafted model: {args.model}", flush=True)
-    model = AutoModelForCausalLM.from_pretrained(
+    # Load as the FULL multimodal Qwen3_5ForConditionalGeneration (NOT AutoModelForCausalLM).
+    # AutoModelForCausalLM resolves qwen3_5 -> Qwen3_5ForCausalLM (text-only), whose weight
+    # keys are flat `model.layers.*` with no vision tower. But vLLM 0.24 only registers
+    # Qwen3_5ForConditionalGeneration, and its hf_to_vllm_mapper expects the checkpoint keyed
+    # `model.language_model.layers.*` (+ `model.visual.*`) — a bare `model.layers.` prefix has
+    # NO mapping rule, so every transformer-layer weight fails to load -> uninitialized weights
+    # -> degenerate `!!!!` output. AutoModelForImageTextToText resolves qwen3_5 ->
+    # Qwen3_5ForConditionalGeneration, so keys are born `model.language_model.*` / `model.visual.*`
+    # matching the working pantheon-27b-mtp-nvfp4 reference. The vision tower loads in BF16 and is
+    # ignored by the quant (re:.*visual.*); calibration is text-only (no pixel_values needed).
+    # (R36 fast-seat namespace fix, 2026-07-14 — the config-merge in the prior recipe was a doomed
+    # patch over a checkpoint quantized in the wrong namespace.)
+    model = AutoModelForImageTextToText.from_pretrained(
         args.model, torch_dtype="auto", device_map="auto", trust_remote_code=True,
     )
     tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)