fix(deepseek): follow pypto-lib 1M-context block tables and bump pypto-lib

[zhangqi-chen]

Follow hw-native-sys/pypto-lib#1220, which compiles DeepSeek V4 Flash MTP for 1048576 positions and splits its MoE into decode and prefill modules.

What this changes

• Submodule 216b497 → 617e165 (pypto-lib#1220 on main). The DeepSeek V4 Flash MTP static logical tables now span 8192 source-token blocks; decode compressed, indexer and compressor-state tables are runtime-width. The other pypto-lib commits between the pins (#1109, #1144, #1179, #1186-#1188, #1190, #1191, #1193, #1195-#1198, #1201-#1203, #1206, #1207, #1216) ride along.
• Block-table widths (npu_runner.py, task_args.py, cli/main.py):
  • prefill ori / cmp / idx tables: 128 → 8192 blocks
  • prefill HCA / CSA / CSA-inner state tables: 2048 / 4096 / 4096 → 131072 / 262144 / 262144
  • decode ori table: 128 → 8192 blocks
  • decode cmp / idx tables stay at 128 blocks; new decode_{hca,csa,csa_inner}_state_table_blocks (2048 / 4096 / 4096, derived from DEEPSEEK_V4_MAX_SEQ_LEN) size the decode state tables, so per-step host metadata keeps its size
  • the max_model_len checks read the decode CSA state table depth, so serving still caps at 16384 until the decode widths are raised
• MoE module eviction: add decode_moe / prefill_moe to the DeepSeek V4 import-context eviction list so each executor load re-specializes them for --ep (moe kept for older pins).

Commits

1. fix(deepseek): follow pypto-lib 1M-context block-table widths
2. fix(deepseek): evict split decode_moe/prefill_moe kernel modules
3. deps: bump pypto-lib to 617e165 for DeepSeek V4 1M context

Verification

• Device (8 cards, DP8/EP8, dsv4-flash-w8a8), serving CI on pypto 747e1e4b: k1-fused completes with the unchanged expected text ("…大典礼的场所…东六宫东") and K=3 passes.
• Known failure: k1-prefix-cache. The 1202-token prompt (max_model_len=2048, 1024-token chunks) fails on the first prefill dispatch on all eight ranks with 507018 / orch_error_code=5 INVALID_ARGS. The bump crosses pypto-lib #1186/#1203 (CP prefill and prefix cache) as well as #1220, and pypto-lib's own serving CI runs only k1-fused, so the responsible lib change is still being isolated.
• Unit: tests/unit/model/deepseek 118 passed; the one failure (test_deepseek_mtp_prefill_reads_only_selected_owner_outputs, FakeWorker copy_from() rejects src_offset) also fails on pristine main. Ruff, header, English, and public-docs checks clean.
• Not run locally: the DSpark 16-card guard. DSpark speculation (K=7): drafter chain, verify, and acceptance #223 noted that pypto-lib past 216b497 (through #1187) deadlocked the DSpark serving dispatch path; this bump crosses that range, so the serving-dspark CI job is the check for it.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: d9a3bb95-d73c-481f-b4ac-9bbf0b0d0e5e

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}