feat: support DSV4 device prefix cache (SWA + C4 + C128).

[Kang-Meng]

feat: support DSV4 device prefix cache (SWA + C4 + C128)

Commit: d54b1af · Branch: feat_dsv4_prefixcache · Repo: xllm
Diff: 30 files, +1242 / −576

Background

DeepSeek-V4-Flash (DSV4) previously could not use the device prefix cache.
master.cpp carried a hard guard that, on detecting a deepseek_v4 model type,
force-set enable_prefix_cache=false and logged a warning. This PR removes that
guard and completes CompositeBlockManager + prefix-cache support for DSV4's
three-segment KV layout (SWA + C4 + C128), so DSV4 can actually hit the
device-side prefix cache.

What changed

1. Removed the DSV4 fallback guard (master.cpp, −11 lines).
   DSV4 now takes the normal composite prefix-cache path instead of being
   silently disabled.

2. LeafCombination classification in CompositeBlockManager
   (largest change; composite_block_manager.cpp +630). The leaf shape is
   classified once at construction and cached; the sequence-level orchestrators
   dispatch on it to pick the trim/mount strategy:

   • FLAT_KV — plain KV (normal / Qwen), no trim
   • FLAT_KV_LINEAR — Qwen3.5-Next GDN (KV + LINEAR restore)
   • SWA_COMPRESSED — DSV4 (SWA + C4 + C128): cross-leaf min,
     C128-stride clamp, SWA tail-continuity, exact-repeat pop
   • UNSUPPORTED — prefix cache off / xtensor

   Adds cache_full_blocks_for_sequence to incrementally insert each leaf's
   newly-forwarded full blocks into its own prefix cache.

3. Per-stride block-hash chains (sequence.h/.cpp).
   DSV4 admission probes SWA/C4/C128 back-to-back each tick with different
   strides (base / 4×base / 128×base). Each stride now keeps its own
   incrementally-extended hash chain, so switching strides no longer discards
   and rebuilds the whole chain. Backed by std::map for pointer stability, so
   a Slice from block_hashes() survives inserts of other strides.

4. LinearStatePrefixCache made gap-tolerant
   (linear_state_prefix_cache.cpp +229). SWA slides out middle window blocks
   and LINEAR deeper hits subsume earlier ones, so match()/insert() tolerate
   middle misses (placeholders): the compute path walks the chain from token 0
   rather than seeding from the previous block, placeholders advance the cursor
   without being stamped/emplaced, and trailing misses are trimmed. The base
   PrefixCache remains solid-prefix.

5. Interface cleanup: dropped the size_t* matched_tokens out-param from
   match() (reach now expressed as returned.size() * block_size_); added an
   LRU-neutral contains().

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.