feat(step3.7): support NextN/MTP heads for speculative decoding

[eauchs]

Summary

Wire up Step-3.7-Flash's num_nextn_predict_layers MTP heads end-to-end so that --spec-type draft-mtp becomes usable on step35 GGUFs. Today the converter silently drops layers ≥ num_hidden_layers and the runtime graph never declares an MTP draft head, so the 3 dense MTP blocks that ship with Step-3.7-Flash sit unused.

This patch follows the existing qwen35moe single-block MTP convention but adapts to Step-3.7 specifics:

• MTP blocks use a dense SwiGLU MLP (mlp.{gate,up,down}_proj), not MoE.
• Per-MTP-block shared LM head (transformer.shared_head.{norm,output}).
• Step35 attention semantics (head-wise sigmoid g_proj, q/k RMSNorm, partial RoPE — n_rot = head_dim / 2 for full-attention).
• Reuses the main tok_embd (Step-3.7 does not ship per-block embed_tokens).

Changes

File	What
conversion/step3.py	Extend block_count by num_nextn_predict_layers; stop filtering HF layers ≥ num_hidden_layers when MTP is enabled; emit step35.nextn_predict_layers; pad per-layer arrays (layer_types, partial_rotary_factors, swiglu_limits[_shared]) so the MTP blocks are marked full-attention with no SwiGLU clamp.
gguf-py/gguf/constants.py	Register NEXTN_{EH_PROJ, ENORM, HNORM, SHARED_HEAD_HEAD, SHARED_HEAD_NORM} on MODEL_ARCH.STEP35.
gguf-py/gguf/tensor_mapping.py	Map Step-3.7's transformer.shared_head.{norm,output} to NEXTN_SHARED_HEAD_{NORM,HEAD}.
src/models/step35.cpp	Read nextn_predict_layers in load_arch_hparams, force trailing blocks to full-attention. Split tensor loading into trunk (MoE + shared expert + Step35 attn) for [0, n_main) and MTP heads (dense SwiGLU + nextn.* + per-block shared head) for [n_main, n_layer). Trim main forward to n_transformer_layers and expose res->t_h_pre_norm so the draft head can seed AR steps. Implement llama_model_step35::graph_mtp with Step35 attention + dense MLP + per-block shared head.
src/models/models.h	Declare llama_model_step35::graph_mtp.

Architecture notes

Step-3.7-Flash ships:

• Main stack: 45 layers (3 leading dense + 42 MoE, iSWA 3:1 pattern, window 512).
• MTP heads: 3 dense blocks at HF layers 45, 46, 47, each with eh_proj, enorm, hnorm, full Step35 attention, dense SwiGLU MLP, and transformer.shared_head.{norm, output}.

The MTP graph implements the DeepSeek-V3 / Qwen3.5 NextN pattern:

h_norm  = RMSNorm_h(h_input)               # nextn.hnorm
e_norm  = RMSNorm_e(embed(prev_token))     # nextn.enorm
x       = nextn.eh_proj(concat(e_norm, h_norm, dim=0))   # [n_embd, n_tokens]
inpSA   = x
x       = input_layernorm(x)               # attn_norm
attn_out= step35_self_attn(x)              # g_proj sigmoid gate + q/k norm + partial RoPE
x       = inpSA + attn_out
x_fres  = x
x       = post_attention_layernorm(x)      # ffn_norm
x       = dense_swiglu_mlp(x)              # mlp.{gate,up,down}_proj
h_next  = x_fres + x
logits  = nextn.shared_head_head(nextn.shared_head_norm(h_next))

h_next is exposed as res->t_h_pre_norm so the AR draft loop can chain MTP steps.

For now graph_mtp only uses the first MTP block (lowest index), matching the existing qwen35moe single-block convention. The driver can iterate the draft graph multiple times manually to exploit all 3 heads; multi-block chained MTP can land in a follow-up.

Test plan

• Build: cmake --build build -j clean on macOS arm64 (Metal + Accelerate + BLAS).
• test-llama-archs passes for step35 on Apple M3 Max / Accelerate / Meta backends.
• Tensor-mapping smoke test: all 51 MTP tensors from the published Step-3.7-Flash safetensors index (3 blocks × 17 tensors) resolve to blk.{45,46,47}.{nextn.*, attn_*, ffn_*} via the updated tensor map.
• End-to-end conversion from the upstream HF Step-3.7-Flash safetensors (requires the original BF16 release; out of reach for this PR author).
• Logit parity vs. StepFun's reference MTP forward (vLLM / SGLang). The HF modeling_step3p7.py does not expose an MTP forward, so the graph here follows the DeepSeek-V3 / Qwen3.5 NextN convention. Maintainer review against your internal reference is the validation that will close the loop.

Notes

• No changes are needed in stepfun-ai/Step-3.7-Flash; the MTP weights are already published in the HF safetensors.
• cache_reuse remains unsupported on step35 (per-layer RoPE dims break K-shift); that's a separate issue.

[eauchs]

Quick heads-up on follow-up validation: I'm capped at 128 GB unified RAM on an M3 Max, which isn't enough to do the next step myself (downloading the ~400 GB BF16 release, re-converting with this patch, and running the model with the 3 MTP heads loaded alongside the main 198B/IQ4_XS stack + KV cache + working set).

So I can confirm the conversion pipeline emits the expected tensor names and that the runtime graph builds + loads against a fresh GGUF, but:

• Re-converting from the upstream HF safetensors
• Loading the resulting GGUF and exercising --spec-type draft-mtp
• Checking logit parity against your internal MTP reference (vLLM / SGLang)

…will need to happen on hardware with enough memory, ideally with someone who has the StepFun reference handy. Happy to iterate on review comments either way.

[forforever73]

My bad — I never enabled notifications for this fork, so I didn't see your PR until now. I've submitted a PR upstream already. Can you try this one and see if it works for you?
ggml-org#24340