Problemsby GlyphOpenAI upgrades break working setups
25 posts from 25 people20d activeSteady, +1 this week
Posts per day
The posts behind it
[Bug]: NixlConnector consumer reads remote KV after the producer lease expired and the blocks were recycled (silent KV corruption, NaN logits, illegal memory access)
### Your current environment - vLLM 0.32.0 (`vllm/vllm-openai:v0.32.0`), torch 2.13.0+cu130, NIXL 1.5.0, UCX (CUDA IPC, both engines on one host) - 2x NVIDIA B200 (driver 580.178.04): one prefill and one decode engine, TP=1 each, `kv_buffer_device=cuda` - Model:...
[BUG] TraceJSONEncoder floods logs with Pydantic serializer warnings for streamed chunks whose fields don't match OpenAI types
### Summary `mlflow.openai.autolog()` records every streamed chat-completions chunk as a span event via `TraceJSONEncoder`. The encoder serializes Pydantic models with `obj.model_dump()`, which keeps Pydantic's default `warnings=True`. When a provider returns a field whose...
[BUG] AI Gateway streaming chat responses never end with `data: [DONE]`
### Issues Policy acknowledgement - [x] I have read and agree to submit bug reports in accordance with the [issues policy](https://www.github.com/mlflow/mlflow/blob/master/ISSUE_POLICY.md) ### Where did you encounter this bug? Local machine (Docker) ### MLflow version - Server:...
bug: reasoning tag parsing stops after the first block; later thinking blocks render as plain text in the answer
...appending to the existing `reasoning_content` item rather than creating another one. ### Workaround A small `stream` filter that pulls every thinking span out of `delta.content` and appends it to `delta.reasoning` works perfectly (state machine over chunk boundaries, ~50 lines)....
Built an open-source recorder that can prove an agent trace wasn’t edited after the fact. Looking for people to break it
...thing too, because the obvious hook in Pydantic AI turned out to be the wrong one.. What doesn't work yet: streaming responses aren't captured (the request is, the response gets marked as not captured), and the signing key sits on the same machine as the traces, which is weaker...
[Bug]: Qwen3.8-27B (GDN hybrid) with PP=3 on sm_86: illegal memory access (Xid 31) with prefix caching, gibberish output without it; regression since v0.28.0 (v0.31.0, v0.32.0, both model runners)
### Your current environment Environment `collect_env` was run with the `vllm/vllm-openai:v0.32.0` image in a pod without GPU access, because all three GPUs were serving. So the CUDA/GPU section below is empty. The host GPU and driver details are listed after it. Host GPU /...
[BUG] AI Gateway guardrails inject a chat base_url into Jev decision judges
> [!WARNING] > Before submitting a PR, please make sure that: > - A maintainer has triaged this issue and applied the `ready` label > - This issue has no assignee > - No duplicate PR exists > > PRs not meeting these requirements may be automatically closed. ### Issues Policy...
[Bug]: compile_mm_encoder is engine-fatal for RADIO/NaViT encoders (Nemotron Parse 3.5)
### Your current environment Stock `vllm/vllm-openai:nightly-83c078d9` (**v0.32.1rc1.dev361+g83c078d95**, contains #61651), 1x H100, `nvidia/Nemotron-3.5-Parse` (`model.safetensors` sha256 `81dd7609…`, the evaluated release). No plugins, no local patches. ### 🐛 Describe the bug...
issue: finished response stays 'generating' (action bar flashes then disappears, next prompt queued) — chat:active handler reloads the chat before the completion handler writes the message back
### Before Submitting - [x] I have searched existing issues and found no report of this (searched "chat:active", "reload", "never shows finished", "action bar disappears"). - [x] Reproduced on the current release; the code paths referenced are unchanged on `main` at the time of...
issue: first message sent right after the page loads disappears or its reply stays empty on a slow connection
Hey! On a slow connection, the first reply in a new chat sometimes stays empty with the Stop button on, until you reload. Not the same as #31966: the page already has the chat id here. What happens is that the new-chat setup from the page load is still running when you hit Send,...
[Bug]: --attention-backend b12x cannot serve hybrid attention/mamba models on v0.32.0 — hybrid page alignment inflates the block size to 2176, B12X only supports (64, 128)
...inflated to 7808 tokens - degrades prefix caching/concurrency rather than hard-failing, because those backends accept larger pages). The in-code `TODO(tdoublep)` notes the constraint can be relaxed by changing the mamba chunk layout. ## Worked on v0.31.0 The identical flag set...
[Bug]: Provider fields on the final usage chunk of a stream are dropped
### Description GreenPT, an OpenAI-compatible provider, returns a top-level `impact` object with the energy and emissions of each request. LiteLLM keeps it on non-streaming responses, but streaming clients never see it, with or without `stream_options.include_usage` When...
[Bug]: AG-UI router: prior tool calls are sent to the LLM as <tool_call> text, so models stop making native tool calls
### Bug Description `ag_ui_message_to_llama_index_message` (utils.py) converts an AG-UI assistant message with `toolCalls` into an assistant message whose *content* is ` … … ` text, and a `ToolMessage` into a `user` message (the code calls this "an opinionated hack for now"). An...
feat: realtime calls should start speaking a hand-off's answer while the chat model is still writing it
...that point is the chat model generating while the call is silent. ### Alternatives or Workarounds - A faster chat model: it shortens the wait but doesn't remove it. - Letting the voice model answer itself: a small voice model makes facts up, which is why we hand off everything...
core: convert_to_openai_messages drops invalid tool calls from exported conversations
### Submission checklist - [x] This is a bug, not a usage question. - [x] I added a clear and descriptive title that summarizes this issue. - [x] I used the GitHub search to find a similar question and didn't find it. - [x] I am sure that this is a bug in LangChain rather than...
[BUG] Bedrock Converse formatter stringifies multimodal system and tool message content
### Description `BedrockCompletion._format_messages_for_converse` (`lib/crewai/src/crewai/llms/providers/bedrock/completion.py`) turns system and tool messages into text with an f-string or `str()`. `LLMMessage.content` can be the parts list (`str | list[dict[str, Any]] |...
[BUG] GenAI semconv export drops OpenAI system messages whose content is a list of text parts
> [!WARNING] > Before submitting a PR, please make sure that: > - A maintainer has triaged this issue and applied the `ready` label > - This issue has no assignee > - No duplicate PR exists > > PRs not meeting these requirements may be automatically closed. ### Issues Policy...
bug: Removing a model from a connection's Model IDs allowlist leaves an orphaned model table row that keeps rendering in Admin → Settings → Models
### Before Submitting - [x] I searched open and closed issues and discussions for an existing report. - [x] I reproduced this bug on the latest release AND on the current `dev` branch, right before submitting this report. I did not just check an old version or rely on a check...
[Bug][SM120] MTP speculative decoding acceptance rate drops to 0% on nightly with native FLASHINFER_MLA_SPARSE_SM120 backend (GLM-5.3-Flash)
## Environment - vLLM: `0.30.1rc1.dev539+g5b6b657e1` (nightly wheel, 2026-10-02 build, commit `5b6b657e1`) - Hardware: 8x RTX PRO 6000 (SM120), PCIe, no NVLink - Model: GLM-5.3-Flash (FP8 checkpoint, TP=8) - flashinfer-python: 0.7.0 - Base image: vllm-openai v0.30.0 + nightly...
[Bug]: Rust vllm-bench openai-chat counts token-less streams as successful
### Your current environment Client-side benchmark issue, not tied to a GPU environment. Reproduced on `main` (macOS, CPU only) against a mock OpenAI-compatible server. ### 🐛 Describe the bug The Rust `vllm-bench` `openai-chat` backend marks a request successful whenever the...
/v1/chat/completions: reordered tool results are associated by position instead of tool_call_id
### What is the issue? ### Summary On Ollama 0.35.1, the OpenAI-compatible `/v1/chat/completions` endpoint can associate parallel tool results by message position rather than by `tool_call_id` on the tested rendering paths. If two tool calls have complete, correctly identified...
Anyone else getting reset tokens that expire right after the weekly reset?
Has anyone else noticed that the reset tokens they receive expire right after their weekly reset? The timing seems odd, since that’s when you’d have the least reason to use them.. Is this happening to anyone else, or is it just how my resets line up?.
Regression: No way to reopen the sidebar from an open chat
[Bug]: Whisper `verbose_json` silently drops the words after the last complete segment at every chunk cut (long audio); `json`/`text` keep them
...the uncovered tail `[t, chunk_end)` as an extra request and append it as a segment. Workaround we use today: cut the audio at silences into ≤ 28 s pieces client-side and send each piece separately (the server never chunks), then offset the segment times. Works, but every...
[BUG] OpenAI Responses instructions stringify multimodal system content as Python repr
### Description OpenAICompletion._prepare_responses_params converts non-string system message content with str(content). LLMMessage.content may be a list of multimodal content parts, so the Responses API instructions receive a Python repr such as [{"type": "text", ...}] instead...
Companies and products named
About this problem
Evidence
25 posts from 25 people in 11 places, about 4 a week over 46 days. Mostly on vllm-project/vllm, open-webui/open-webui, mlflow/mlflow. Tools named alongside: vLLM, GitHub, NVIDIA, Docker.
Frustration Frustration 1 of 3· Seen on GitHub issues, Reddit, Community forums
How it was grouped
Posts that state a pain and match the "breaking-changes" rule about OpenAI. First post Sep 21, 2026, latest Nov 5, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
The same problem elsewhere
GitHub upgrades break working setups
87 posts from 75 people on GitHub issues, Hacker News, Stack Overflow.
Upgrades and new releases break working setups
76 posts from 62 people on GitHub issues, Reddit, App Store reviews.
Claude Code upgrades break working setups
68 posts from 61 people on GitHub issues, Hacker News, Reddit.
Codex upgrades break working setups
65 posts from 58 people on GitHub issues, Reddit.
DuckDB upgrades break working setups
37 posts from 28 people on GitHub issues.
Claude upgrades break working setups
35 posts from 34 people on GitHub issues, Reddit, Hacker News.
ClickHouse upgrades break working setups
31 posts from 20 people on GitHub issues.
Anthropic upgrades break working setups
27 posts from 22 people on GitHub issues.
PyTorch upgrades break working setups
26 posts from 21 people on GitHub issues, Reddit.
ChatGPT upgrades break working setups
26 posts from 26 people on GitHub issues, Community forums, App Store reviews.
Open WebUI upgrades break working setups
25 posts from 2 people on GitHub issues.
Gemini upgrades break working setups
24 posts from 24 people on GitHub issues, Community forums, Reddit.
Other OpenAI problems
Trouble with OpenAI responses
Top post: "completion_cost() raises ModelNotMappedError for groq/openai/gpt-oss-* responses, even when model=groq/... is", one of 18 posts from 16 people on GitHub issues, Community forums.
OpenAI users keep asking for help
Top post: "Multi-file uploads failing and ZIP archives not parsing / sandbox timeout (100% Limits Remaining)", one of 47 posts from 45 people on Community forums, GitHub issues, Reddit.
Trouble with OpenAI assistant
Top post: "Bedrock Converse forwards OpenAI-style content parts in user and assistant messages unconverted", one of 3 posts from 3 people on GitHub issues.
OpenAI fails to return valid structured output
12 posts from 12 people on GitHub issues, Hacker News.
OpenAI runs out of context on real work
8 posts from 8 people on GitHub issues, Community forums.
OpenAI rate limits and usage caps block work
6 posts from 6 people on GitHub issues, Community forums.
Trouble with OpenAI blocked
Top post: "issue(a11y): pinch zoom is blocked on phones because the viewport sets maximum-scale=1", one of 3 posts from 3 people on Community forums, GitHub issues.
Trouble with OpenAI calls
Top post: "Unable to connect to dot calls across mobile, web and desktop", one of 7 posts from 6 people on GitHub issues, Community forums.
Running OpenAI locally hits memory and speed limits
13 posts from 13 people on GitHub issues.
Trouble with OpenAI chat
Top post: "Chat attachments inside Projects intermittently fail to materialize into /mnt/data, while File Library and", one of 12 posts from 12 people on Community forums, GitHub issues, Reddit.
Trouble with OpenAI image
Top post: "Image generation is broken for serious work", one of 8 posts from 8 people on GitHub issues, Community forums, Reddit.
Trouble with OpenAI decisions
Top post: "OpenAI Decisions API", one of 6 posts from 4 people on GitHub issues.
History