Issues

• Host RAM (Private bytes) of flm serve grows with every new context length and is never released (v1.0.11, Windows HRX)

  ROCm/FastFlowLM#903 · Filter by author LalaDango opened on Oct 31, 2026

• Gemma4 tool-call parsing splits a call when a string argument contains a double quote (v1.0.11)

  ROCm/FastFlowLM#900 · Filter by author markl-a opened on Oct 31, 2026
  

• [Issue]: qwen3.5:9b: after structured-output (JSON schema) requests the server gets stuck producing "//////" for every request until restart

  ROCm/FastFlowLM#893 · Filter by author Matzebhv opened on Oct 30, 2026
  

• Server: min_p is read as int, so fractional values (e.g. 0.05) become 0

  ROCm/FastFlowLM#887 · Filter by author LalaDango opened on Oct 30, 2026
  

• gemma4-it:12b: with thinking off and tools, the tool-call opener is sampled as <|channel> instead of <|tool_call>, so calls leak into content or loop until max_tokens

  ROCm/FastFlowLM#886 · Filter by author LalaDango opened on Oct 30, 2026
  

• [Bug]: d3 decision requests keep ~13 CPU cores busy (thread pool spin-waits); no way to cap threads

  ROCm/FastFlowLM#882 · Filter by author chetan-metrum-ai opened on Oct 30, 2026
  2

  

• [Bug]: ignore_eos and min_tokens are not honored by /v1/chat/completions

  ROCm/FastFlowLM#881 · Filter by author chetan-metrum-ai opened on Oct 30, 2026
  

• Status of the official gguf → q4nx conversion tool announced in v0.9.26 — and supported path for fine-tuned Qwen3/Qwen3.5?

  ROCm/FastFlowLM#828 · Filter by author ScottyVenable opened on Oct 22, 2026
  1

• Greedy decoding (top_k=1) is not reproducible on NPU — identical request gives different output

  ROCm/FastFlowLM#810 · Filter by author Influx934Threads opened on Oct 18, 2026
  1

• Host RAM: keep lookup-only tables (token and per-layer embeddings) quantized and memory-mapped instead of resident copies (saves ~1 GB on Qwen3.5-2B, ~2.9 GB on Gemma 4 E2B)

  ROCm/FastFlowLM#807 · Filter by author Jcollier0120 opened on Oct 16, 2026
  1

• libwhisper_npu leaks 1.33 MB per encode_audio; NPU memory is only returned by a full HRX shutdown

  ROCm/FastFlowLM#792 · Filter by author quine00 opened on Oct 15, 2026
  1

  

• [Issue]: Vectors collapse (non-discriminative) with the xclbins shipped for ≥0.9.45

  ROCm/FastFlowLM#789 · Filter by author fermosan opened on Oct 14, 2026

• whisper_npu: WhisperEncoder::add_bias_optimized leaks 1,331,200 B on every encode_audio() call

  ROCm/FastFlowLM#788 · Filter by author superm1 opened on Oct 14, 2026

• [Issue]: HRX preview (v1.0.8, Linux): flm bench gemma4-it:e2b fails at 32k with 'ert state 8'; XRT completes, serve works

  ROCm/FastFlowLM#785 · Filter by author mariolpantunes opened on Oct 13, 2026
  1

• [Issue]: Model weigths are being downloaded to XDG_CONFIG_HOME

  ROCm/FastFlowLM#782 · Filter by author christian-heusel opened on Oct 12, 2026

• [Feature]: flm list sort by date

  ROCm/FastFlowLM#772 · Filter by author gdkrmr opened on Oct 9, 2026

• [Feature]: Support for Google's EmbeddingGemma 2

  ROCm/FastFlowLM#771 · Filter by author BoneHorror opened on Oct 6, 2026
  1

• [Feature]: Quantized version of Qwen3.8-27B

  ROCm/FastFlowLM#769 · Filter by author Skurios opened on Oct 6, 2026
  6

• [Issue]: fail to load Qwen3.8-mtp:27b, flm version 1.0.7

  ROCm/FastFlowLM#767 · Filter by author tangmeng09 opened on Oct 4, 2026
  12

• flm serve: a client that resets before accept stops the server accepting connections (HttpSession ctor throws ENOTCONN)

  ROCm/FastFlowLM#762 · Filter by author noamsto opened on Sep 28, 2026
  1

• flm serve error responses: sent with HTTP 200, include raw exception text, and GET /api/ps is undefined behaviour with no model loaded

  ROCm/FastFlowLM#761 · Filter by author noamsto opened on Sep 28, 2026

• flm serve: /v1/completions, /api/generate and /api/chat keep decoding after the client disconnects

  ROCm/FastFlowLM#760 · Filter by author noamsto opened on Sep 28, 2026
  1

• flm serve: /v1/completions and other routes run outside the NPU queue and race with model swaps

  ROCm/FastFlowLM#759 · Filter by author noamsto opened on Sep 28, 2026
  1

• /v1/embeddings ignores prompt_name / task_type, so every input is embedded as a query

  ROCm/FastFlowLM#758 · Filter by author noamsto opened on Sep 28, 2026

• flm serve: streaming /api/chat segfaults the server on every request (insert() twice, generate() never called)

  ROCm/FastFlowLM#757 · Filter by author noamsto opened on Sep 28, 2026