llama.cpp engine returns no per-token logprobs in the streaming path (all GGUF models)

[ttupper92618]

Summary

The llama.cpp (GGUF / AMD-Vulkan, kite4) engine does not emit per-token logprobs in the streaming chat path. Every GGUF model in the battery failed the logprobs-parity test with top_logprobs requested, while the same request shape works on the MLX engine.

Evidence (live, dev, 2026-06-23 battery)

logprobs-parity failed on all 5 GGUF models across both llama.cpp cells:

• gguf-llama-cpp: Llama-3.2-1B-Instruct-GGUF
• gguf-big: Qwen2.5-7B-Instruct-GGUF, Qwen3-Coder-30B-A3B-Instruct-GGUF, Llama-3.3-70B-Instruct-GGUF, openai_gpt-oss-120b-GGUF

Each: Expected per-token logprobs but the stream returned none, output empty for that test. The other tests (concise-factual, ordered-integers, tool-call-path) passed on the same models, so generation and tool-calling are fine; only logprobs are missing.

Scope / likely cause

logprobs was part of the AMD v1 llama.cpp capability-parity work, so this is probably a streaming-specific gap: the non-streaming path may attach logprobs while the SSE/streaming path drops them (or logprobs=true / top_logprobs is not being threaded into the llama-cpp-python streaming call, or the per-chunk choices[].logprobs.content is not populated). The harness reads logprobs at the OpenAI-standard choices[0].logprobs.content location, which MLX populates.

Suggested next step

Confirm whether non-streaming /bench/chat/completions returns logprobs for a GGUF model; if it does and streaming does not, fix the streaming emitter to carry per-token logprobs (and top_logprobs alternatives) the same way MLX does.

Impact

Any logprobs-dependent client feature (token confidence, MTP acceptance analysis, structured-output scoring) is unavailable on AMD/llama.cpp nodes, silently returning no logprobs rather than erroring.

[ttupper92618]

Fixed by #397 (merged to dev). A logprobs request to a llama.cpp node without logits_all now fails with a clear, actionable error (naming SKULK_LLAMA_CPP_LOGITS_ALL=1) delivered as an error chunk, instead of silently returning empty logprobs. top_logprobs is normalized to a logprobs request once at the chat-completions boundary so MLX and llama.cpp behave consistently (explicit logprobs wins; inferred only when unset). Enabling logprobs stays opt-in per node. Tests cover the runner guard, kwargs translation, and adapter normalization.