dsv4.1: Engram module and request history support

[hnyls2002]

Summary

• Add the DeepSeek-V4.1 Engram module (layers/engram.py): hashed n-gram lookup into per-layer fp8 tables with e8m0 block scales, gated into the hc residual stream. Hashing, table gather and the gate run on the existing kernels/ops/embeddings/engram_* Triton kernels; the torch paths are the ROCm/CPU fallbacks.
• Give NgramEmbeddingManager the request-history bookkeeping the hasher needs: a per-request history row seeded on extend (ScheduleBatch.engram_history -> ForwardBatch.engram_history), a PD-decode seed for prebuilt requests, and update_after_verify for speculative verify commits (the caller lands with the DSpark wiring).
• ModelConfig: use_engram (from engram_layer_ids) and ngram_embedding_n next to the existing use_ngram_embedding.

Changes

Hash layout

• EngramLayout.from_config draws one ascending prime sequence in (layer, n-gram size, head) order; EngramHasher maps token ids through a tokenizer-normalized compressed vocabulary and keeps a [req_slots, n - 1] history table. Primality is a local deterministic Miller-Rabin, so no new dependency.

Table placement

• Default: rows sharded over the TP group in device memory, reassembled with an all-reduce; under DP attention the lookup gathers every DP rank's ids first.
• SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1 keeps the tables in pinned host memory and gathers rows over the CPU link. SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT selects shared (default: one buffer per TP group, no lookup all-reduce; the ranks must share a PID namespace) or per_rank (one shard per rank, all-reduce kept, huge pages via MADV_COLLAPSE after dropping the checkpoint page cache). The mapping lives for the whole process.

Shared detectors and config

• NgramEmbeddingManager keeps its LongCat token-table behavior unchanged for models without engram_layer_ids; the unused k field is dropped (no readers).

Verification

• The compressed vocabulary size the tokenizer normalizes to is asserted against engram_compressed_vocab_size when the hasher is built, so a tokenizers behavior change fails at startup instead of silently rehashing.
• Existing manager, scheduler and config CPU suites pass unchanged; the MLX runner stubs gain the two new ModelConfig fields.
• Host-table smoke: shared and per_rank tables construct, register and report huge-page coverage; unknown layout values fail at startup.

---

CI States

Latest PR Test (Base): 🚫 Run #35185092582
Latest PR Test (Extra): ❌ Run #35190590426
Latest PR Test (AMD ROCm 10): 🚫 Run #35185092572

[hnyls2002]

/rerun-test test_ngram_embedding_manager.py test_scheduler_hicache_events.py test_scheduler_chunked_req_gate.py test_auxiliary_output.py

[github-actions]

Results for /rerun-test test_ngram_embedding_manager.py test_scheduler_hicache_events.py test_scheduler_chunked_req_gate.py test_auxiliary_output.py:

🚀 ubuntu-latest (4 tests): ✅ View workflow run

cd test/ && python3 registered/unit/model_executor/model_runner_components/test_ngram_embedding_manager.py
cd test/ && python3 registered/unit/managers/test_scheduler_hicache_events.py
cd test/ && python3 registered/unit/managers/test_scheduler_chunked_req_gate.py
cd test/ && python3 registered/unit/managers/test_auxiliary_output.py

[hnyls2002]

/tag-and-rerun-ci