[ddvnguyen]

Summary

llama_state_seq_set_data returns the number of bytes read on success and 0 on failure — per its own doc comment in include/llama.h:

// Returns:
//  - Positive: Ok
//  - Zero: Failed to load

server-context.cpp:4534's original check was if (status != 0), backwards — it treated every successful restore (nonzero = bytes read) as a failure. Confirmed against upstream ggml-org/llama.cpp directly (not just this fork's copy of the doc comment): src/llama-context.cpp's implementation is untouched by the fork, and every real call site upstream (tools/server/server-task.cpp, tests/test-save-load-state.cpp) checks the expected byte count, never != 0 as failure.

Correction during review: the first fix pushed here changed the check to if (status != kv_len). That is also wrong, for a subtler reason: status (llama_state_seq_set_data's return value) only counts bytes the KV-cache reader itself consumed (llama_context::state_seq_read_data returns io.n_bytes(), a running counter of actual read()/read_tensor() calls — it has no knowledge of the buffer size passed in beyond an overrun guard). kv_len here is KV-state bytes plus trailing logits: PREFILL_DONE unconditionally appends n_vocab * sizeof(float) logit bytes after the KV state whenever logits are available (~line 4098) — the normal case for merged-decode, not an edge case — and the coordinator forwards that full blob verbatim as the DECODE 0x43 payload. So status < kv_len is the expected, correct outcome whenever logits are present, and comparing against kv_len reintroduces a false failure on essentially every real restore.

Fixed to if (status == 0), matching the STATE_PUT sibling check exactly (server-context.cpp ~line 3395).

Also fixed: DECODE_APPLY was missing the trailing-logits extraction entirely. The STATE_PUT handler computes remaining = state_len - n_read and, when it equals n_vocab * sizeof(float), injects those floats into the per-slot restored_logits buffer so the first sample after restore doesn't read garbage (server-context.cpp ~line 3405-3417, comment: "without these, common_sampler_sample reads garbage after StatePut"). DECODE_APPLY restored the KV cache but silently dropped this step. Ported the same logic in.

Evidence this is the real bug behind #78

157626320 in #78's reported log line (KV restore failed (157626320)) is ≈150.3 MiB — matching almost exactly "the 150 MB KV blob" the issue itself describes as the input being restored. The restore was succeeding the whole time; the check just reported it as failed. Same shape independently observed in an earlier debug session (KV restore failed (36656)).

Provenance

The inverted check was introduced fresh in PR #64 ("feat(rpc): ggml-org#470 Phase 2 — framed DECODE (0x43) rewrite", merged 2026-07-24) and survived two subsequent PRs that touched adjacent lines for unrelated reasons — #66 (2026-07-28) and #69 (2026-07-29, Closes ggml-org#506) — without the polarity being caught in any of those review rounds. The missing logits-injection step also dates to the same PR #64 (DECODE_APPLY was written without porting the full STATE_PUT success path).

Verification

Not yet build- or live-rig-verified — no local build directory was available in this worktree. The changes are localized to the DECODE_APPLY success/failure branch, use only existing fields (slot->restored_logits, slot->logits_valid, vocab) already used identically in the STATE_PUT sibling a few hundred lines above, and introduce no new symbols. Live-rig E2E verification (the project's own required step before merge) is still outstanding.

Closes #78