[void-mckenzie]

feat(gguf): load mtp drafts for the qwen3.5 family from nextn blocks and mtp-*.gguf sidecars

The GGUF side branch of #828. #844 gave the Qwen3.5 family MTP from safetensors; GGUF targets were the only ones still refused. This adds only weight loading, the verify path is #844's unchanged.

• The draft reads the trunk's own nextn block, the block llama.cpp keeps after the last decoder layer (block_count includes it, {arch}.nextn_predict_layers says how many). Exactly one nextn layer is accepted; a --no-mtp conversion or a Strata-style MTP pack is refused with a message that names both.
• The block's tensors land under the draft's state-dict names: nextn.* as fc and norms, the block itself as layers.0, q output-gate rows packed in attn_q, gate|up fused. The draft's quant comes from the file's own tensor table, so each ggml mix loads as stored; the MoE draft's routed experts load as its one resident bank set.
• --speculative-draft-model-path also takes an mtp-*.gguf sidecar (unsloth's MTP/ releases). Those files keep the trunk's full block_count KVs so blk.N addressing matches; a sidecar is accepted only when it holds exactly the one nextn block, the target's own architecture and hidden size, otherwise the refusal says what the draft of a GGUF target is. A --no-mtp trunk plus a sidecar works: the trunk config's shapes, the sidecar's weights.
• docs/cli.md speculative section updated to match.

Verified

• pytest tests/models/test_qwen3_5_moe_weight.py tests/models/test_models_registry.py tests/spec/: 67 passed, 104 skipped (skips are pre-existing hardware/checkpoint gates). The new tests cover the block-to-state-dict mapping, quantized packed leaves, resident expert banks, the no-MTP refusal message, and the sidecar draft loading under its own quant. There is no numeric parity test against the HF mtp.* draft of the same checkpoint yet; the mapping is checked tensor by tensor against a synthetic file instead.
• Live serving, GPU 0 (RTX 5090), greedy:
  • Dense unsloth/Qwen3.8-27B-GGUF UD-Q5_K_XL, ft serve ... --speculative-algorithm mtp --num-tokens 24576: 191.3 t/s mean steady-state decode at accept len 3.88-4.00 (k=3, at the theoretical max); the same boot without spec runs 73.2 t/s, 2.61x. Both boots answer "What is 15% of 240?" with byte-identical content and reasoning fields ("36"), and a 2048-token long-form prompt is also byte-identical between them.
  • MoE Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf (trunk's own nextn): 442-486 t/s, accept len 3.65-3.85; the draft's nextn experts resident on GPU (0.5 GiB) beside the trunk's 40 offloaded layers.
  • Sidecar: the same 27B trunk with --speculative-draft-model-path .../MTP/mtp-Qwen3.8-27B-Q4_0.gguf: greedy 17*23+45/9 answers 396, accept len 2.0-3.5, 113 t/s steady state.

[jason-fxz]

Tested d1b3690 against its base 9cacd08 on real GGUF files; LGTM.

Environment: NVIDIA H100 80GB HBM3, driver 580.95.05, CUDA 13.0, Intel Xeon Platinum 8480+ (26 vCPU), torch 2.14.1+cu130, transformers 5.18.0.

Checkpoints:

• unsloth/Qwen3.5-2B-MTP-GGUF BF16 and Q8_0 (trunk with its nextn block), unsloth/Qwen3.5-2B-GGUF Q8_0 (no nextn)
• unsloth/Qwen3.8-27B-GGUF UD-Q4_K_XL and MTP/mtp-Qwen3.8-27B-Q4_0.gguf
• unsloth/Qwen3.6-35B-A3B-MTP-GGUF UD-Q4_K_XL
• bf16 sidecars converted from Qwen/Qwen3.5-2B and Qwen/Qwen3.6-35B-A3B with llama.cpp e937b17a2 convert_hf_to_gguf.py --mtp --outtype bf16
• HF references: Qwen/Qwen3.5-2B, Qwen/Qwen3.6-35B-A3B (run with moe_strategy="offload", moe_cache_auto=True), Qwen3.8-27B, all bf16

Workload: AIME25 problem 1 through the offline LLM API with max_running_req=4, speculative_algorithm="mtp" (and speculative_draft_model_path for sidecars). Greedy runs end naturally (cap 16384) and are compared token by token; speed runs use temperature 1.0, top_p 0.95, top_k 20, natural end, 5 runs, median.

Greedy, token by token:

• spec off, PR vs base: identical on 2B BF16, 2B Q8_0, 2B no-nextn Q8_0, 27B UD-Q4_K_XL and 35B UD-Q4_K_XL (fused).
• HF Qwen3.5-2B with MTP, PR vs base: identical.
• 35B fused vs offload with moe_cache_rate=0.25: identical with spec off and with MTP.
• own nextn vs sidecar draft: identical output on 2B, 27B and 35B.
• 2B no-nextn trunk with --spec-algo mtp and no sidecar: refused at startup with the nextn ValueError.

Sampled speed (tok/s, median of 5; tokens per step in brackets):

target	draft	spec off	MTP	HF bf16 MTP, same model
2B Q8_0	own nextn	431.7	818.7 (3.34)	753.0 (3.23)
2B no-nextn Q8_0	bf16 sidecar		785.3 (3.35)
27B UD-Q4_K_XL	own nextn	86.7	161.0 (3.17)	123.9 (3.10)
27B UD-Q4_K_XL	mtp-Q4_0 sidecar		163.4 (3.20)
35B UD-Q4_K_XL, fused	own nextn	208.7	428.6 (3.22)	360.6 (3.25)
35B UD-Q4_K_XL, fused	bf16 sidecar		427.8 (3.23)
35B UD-Q4_K_XL, offload 25%	own nextn		123.8 (3.22)
35B UD-Q4_K_XL, hybrid 25% (1024 tokens, 3 runs)	own nextn	127.5	190.9 (3.44)

ft serve with the 27B trunk, --spec-algo mtp, with and without --spec-draft-model .../MTP/mtp-Qwen3.8-27B-Q4_0.gguf: both answer a chat request correctly, and only the sidecar run maps the sidecar file.

ft serve --model-path Qwen3.8-27B-UD-Q4_K_XL.gguf --spec-algo mtp
ft serve --model-path Qwen3.8-27B-UD-Q4_K_XL.gguf --spec-algo mtp --spec-draft-model MTP/mtp-Qwen3.8-27B-Q4_0.gguf

Not covered: a --no-mtp MoE trunk with a sidecar, prism or NVFP4 sidecars, ROCm.