Repository navigation
SafeTensors serve route drops ignore_eos and seed: apr-cli's ChatCompletionRequest has no such fields #2853
Description
Found by the cross-vendor review arm (agy, gemini-3.1-pro-high) on PR #2851's third commit, re-verified by reading the code.
crates/apr-cli/src/commands/serve/types.rs ChatCompletionRequest carries no ignore_eos (nor seed), and parse_chat_completion_request in serve/chat.rs deserialises straight into it, so both are silently dropped on the routes that use it: safetensors_chat_completions_handler and the Ollama handlers mounted in serve/safetensors.rs (apr serve run model.safetensors). A client asking for ignore_eos: true there gets a generation that stops at EOS with no refusal — the silent-drop shape PERF-039 closed on the GGUF/realizar route (aprender-serve/src/api/cuda_chat_backend.rs fails closed with reject_unsupported_ignore_eos, and stop_tokens_unless_ignore_eos empties the stop set).
Not on the measured path: the PP-26 witness and the band producer drive the GGUF route, where ignore_eos is honoured (every slot in evidence/perf041/lambda/witness.json completes exactly n_predict tokens). #2851 touches serve/types.rs only to name DEFAULT_CONTEXT_LENGTH (PP-14); the request type's fields are as they were on origin/main, and chat.rs / safetensors.rs are untouched. Fix: add the two fields to the type and either honour them in the SafeTensors decode loop or refuse them the way PERF-039 does.
🤖 Generated with Claude Code
Activity
noahgift commentedon Sep 23, 2026
milestone → 0.72.0 per docs/roadmaps/milestone-triage-2026-09-23.md (#4024), operator approval 2026-09-23 (relayed by aprender-cf). Reason: serve/agent/telemetry surface: Agent Ready (#4000)
noahgift commentedon Sep 25, 2026
Parked by APR-EPIC-001: over S3 budget (rank 6/142). Reopen = pull. Applied by aprender-d1 under a 3-lane quorum receipt (handoff/balance-quorum-rev4.json).
3 remaining items
noahgift commentedon Sep 25, 2026
Reopened by APR-EPIC-001 retro quorum: the agy lane voted keep-open (P0 on its train), and a close needs both counted lanes (cop ruling 2026-09-25, handoff/park-retro-quorum.json). Unlinked; milestone restored to its train. Applied by aprender-d1.
noahgift commentedon Oct 14, 2026
P0 to P2 (2026-10-14 P0 triage): P0 now means it blocks the next release or carries the operator's own words on it; this issue does neither, and its milestone is 0.72.0.
noahgift commentedon Oct 14, 2026
Folded into #2582 as a checklist row (TRIAGE-100, 2026-10-14): the open-ticket cap is 100, so this ticket is now tracked on that master and closed as not planned. Nothing is lost: reopen it when its work resumes, and an open PR for it reopens it. Pushed branch:
fix/2853-safetensors-ignore-eos,fix/2853-safetensors-ignore-eos-seed.Agent: aprender-clerk