[ddvnguyen]

Closes ggml-org#506

Root Cause

In the DECODE_APPLY handler (async phase of the DECODE 0x43 path), slot->prompt_clear(false) clears slot->prompt.tokens before loading KV state via llama_state_seq_set_data. After the KV restore, the token list is never restored.

When the subsequent COMPLETION task runs update_slots():

1. slot.prompt.tokens is empty (cleared, never restored)
2. The n_common decision compares empty stored tokens vs N incoming tokens → n_common = 0
3. n_past = 0 (else branch of n_common decision)
4. pos_next = slot.prompt.tokens.pos_next(0) = 0
5. seq_rm(ctx_tgt, slot.id, 0, -1) — wipes the entire KV cache that was just loaded
6. Slot stalls at n_past=0, n_decoded=0 for the full timeout window

The STATE_PUT handler (old HTTP-completion path) correctly parses the v2 blob header and restores tokens (lines 3357-3362). The DECODE_APPLY handler did not.

Fix

Parse the v2 blob header in DECODE_APPLY to:

1. Extract the token list from the blob header and restore slot->prompt.tokens
2. Skip past the v2 header before passing raw KV data to llama_state_seq_set_data
3. Set slot->just_restored = true so the CHECKPOINT_SEARCH path uses the cached n_past

The fix is backward-compatible: if the coordinator sends raw KV data (no v2 header), the code falls through to the existing behavior.

Evidence

Clean re-test (2026-07-29, build db1654c):

• restore_kv_ms=571 — STATE_PUT/DECODE RPC completed successfully
• P100 slot: n_past=0, n_decoded=0 for 291 seconds until timeout
• No crash, no error logged — the KV was loaded then silently destroyed by seq_rm

Build

CUDA 13.2 sm_86+sm_120 build: 480/480 targets passed.

[ddvnguyen]

Reviewed the diff and root-cause explanation — this checks out. DECODE_APPLY's prompt_clear(false) wiping slot->prompt.tokens before the v2-blob-aware KV restore, leaving n_common's decision logic with nothing to match against, is a precise explanation for the exact symptom in ggml-org#506 (restore RPC succeeds, n_past stuck at 0 for the full timeout, no crash). The fix correctly detects the v2 blob (same 0x02 magic byte check already used in the STATE_PUT handler at server-context.cpp:3276), restores the token list, skips the optional checkpoint sub-blob before advancing past the header, and sets just_restored so the downstream CHECKPOINT_SEARCH path uses the cached n_past instead of recomputing from an empty token list.

One non-blocking note: this duplicates the v2-header-parsing logic that already exists inline in the STATE_PUT handler (~line 3273-3300) rather than factoring it into a shared helper. Not a regression — STATE_PUT's version is also inline, so this follows existing style rather than introducing a new pattern — but worth a follow-up cleanup given there are now two independent inline copies of this parsing logic that could drift out of sync.

Build passes per the PR. No blocking concerns — good catch and a well-targeted fix.