Add Multi-Token Prediction (MTP) inference support

[ArthurZucker]

Summary

• Wires MTP speculative decoding into generate() for DeepSeek-V3 and GLM-4 MoE, loading the MTP modules that ship in the original checkpoints (previously ignored via _keys_to_ignore_on_load_unexpected).
• New GenerationConfig.use_mtp=True opt-in (default False; no behavior change when off). Routes to a new GenerationMode.MTP_DECODING / _mtp_decoding that uses the model's own MTP heads as the draft, verifies in a single forward, reuses the existing _speculative_sampling helper.
• Adds a tiny MTPLayer (RMSNorm + concat-and-project + decoder block + shared_head) mirroring vLLM's deepseek_mtp.py. Base forward keeps iterating only self.layers[:num_hidden_layers]; MTP is reached via a new model.forward_mtp(...) helper.
• _assisted_decoding / candidate_generator.py are untouched.

Scope

• generate(..., use_mtp=True) is supported for DeepSeek-V3 / GLM-4 MoE at batch_size=1 with dynamic cache.
• generate_batch with use_mtp=True raises NotImplementedError for now — continuous batching needs paged-cache slot reservation (K+1 per MTP request) plus per-request accept/reject in _forward_process_and_sample. That will come as a follow-up PR so it gets a clean review on its own.
• Training-time MTP loss is explicitly out of scope.

Design decisions

Draft	Model's own MTP heads (shared weights + shared KV cache).
Path	Dedicated _mtp_decoding, not routed through _assisted_decoding.
Activation	Explicit use_mtp=True; depth = config.num_nextn_predict_layers.
Layout	Tail of self.layers holds MTP modules (matches the checkpoint key scheme model.layers.{num_base + k}.*).

Files

• Generation: generation/configuration_utils.py (new use_mtp flag + mode), generation/utils.py (new _mtp_decoding, mode registration, validation), generation/continuous_batching/continuous_api.py (explicit refusal for now).
• Models: deepseek_v3/ and glm4_moe/ modular/config/modeling. Downstream variants (glm4_moe_lite, glm4v_moe, glm_moe_dsa, longcat_flash, solar_open, youtu) got regenerated with the new field propagated through the modular chain.
• Tests: new tests/generation/test_mtp.py.

Test plan

• pytest tests/generation/test_mtp.py — 9/9 passing:
  • use_mtp=True routes to GenerationMode.MTP_DECODING.
  • Greedy output matches plain _sample token-for-token for K=1/2/3 on DeepSeek-V3 and GLM-4 MoE (random-init, so drafts usually miss → verified bonus-token fallback path).
  • use_mtp=True with num_nextn_predict_layers == 0 raises a clear ValueError.
  • Base forward output is unchanged when MTP layers are added (shared-weight copy + compare).
  • model.forward_mtp produces (hidden, logits) of the expected shapes.
  • ContinuousBatchingManager(..., GenerationConfig(use_mtp=True)) raises NotImplementedError.
• pytest tests/generation/test_candidate_generator.py tests/generation/test_configuration_utils.py — unchanged / green (regression check).
• Import smoke-test on every modular-regenerated downstream model (glm4_moe_lite, glm4v_moe, glm_moe_dsa, longcat_flash, solar_open, youtu).
• make style clean.
• make fix-repo clean apart from a pre-existing mlinter._using_rule_specs / check_modeling_structure.py --rules-toml toolchain mismatch that also fails on an unmodified checkout — unrelated to this PR.

Follow-ups

• generate_batch + MTP (paged-cache reservation, per-request accept/reject).
• Prefill through the MTP layers during the initial forward, to seed their KV caches for better draft quality (currently each step's MTP self-attn only sees the current query token; degraded acceptance vs vLLM but correct).
• Real-weight end-to-end test on deepseek-ai/DeepSeek-V3 once GPU capacity is available.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: deepseek_v3, glm4_moe, glm4_moe_lite, glm4v_moe, glm_moe_dsa, longcat_flash, solar_open, youtu

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

Superseded by a cleaner refactor (MTP moved out of the model into MTPCandidateGenerator). Reopening from the origin branch.