- 24 Jun, 2026 16 commits
-
-
Stefy Lanza (nextime / spora ) authoredllama.cpp's CUDA backend and PyTorch sharing one process corrupt the CUDA context: after a GGUF model runs on an NVIDIA card, the next torch kernel dies with "CUDA error: invalid argument". Seen in a township match — a gemma GGUF served a chat completion on the nvidia engine, then the next Z-Image (diffusers) image gen crashed in its Qwen3 text-encoder's sdpa mask (padding_mask.all()). Every Z-Image run before llama.cpp touched CUDA succeeded; the first one after crashed. Evict+swap in one process can't fix it — torch holds the (now-corrupted) context for the process lifetime. Fix: when GPUs are auto-detected and server.isolate_gguf_engine is True (default), each NVIDIA torch engine gets a co-located sibling gguf engine on the SAME card (own process -> own CUDA context). The torch engine drops the `gguf` capability and serves transformers/diffusers; the sibling (backend=nvidia so GGUF takes the proven CUDA-llama path, capabilities={gguf}) serves llama.cpp. Routing is already capability-based, so GGUF goes to the gguf engine and HF/diffusers to the torch engine. Both are real engine subprocesses, so the front's routing, VRAM/eviction and thermal (cooperative pause + SIGSTOP on the process group) apply unchanged — and SIGSTOPping the gguf engine no longer freezes the torch engine. Ignored when engine_specs is set (declare the split yourself). Disable with server.isolate_gguf_engine=false. Design + trade-offs (extra torch context, no cross-engine eviction on a shared card) in docs/gguf-process-isolation.md. Co-Authored-By:Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RdMufYvtTbtGDWsiZVoXce
-
Stefy Lanza (nextime / spora ) authored
Trained LoRAs, characters, environments and voices are resolved from the legacy HOME-based dir ~/.coderai, not from --config. In the container the user's home (e.g. /home/ubuntu) isn't where config is mounted ($CONFIG_DIR/coderai), so those lookups missed the mounted data and image-gen failed with "LoRA '<name>' not found on server". Symlink ~/.coderai -> $CODERAI_CONFIG_DIR/coderai so HOME-based and --config lookups agree, removing the need for an extra per-deployment --map. Acts only when safe (a symlink, missing, or an empty dir) so a user-mounted volume or real data is never clobbered. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The models page reads /admin/api/cached-models (served by the system worker), whose _scan_caches() merges each model's saved config (precision/quant/ctx) into the cards. The scan freshness check _scan_signature() only fingerprinted the HF/GGUF cache directories, never models.json, and model-configure never invalidated the scan. A config save rewrites models.json but touches no cache file, so the signature stayed unchanged and the scan was served stale for the full TTL (600s) -- the engine reloaded but the front showed old config until a full restart. - _scan_signature(): fold models.json mtime into the signature so any process serving cached-models re-scans on the next read after a save. - system worker /internal/reload-config: call _invalidate_cache_scan() after rebinding config, so the front's reload-push refreshes the scan immediately. - engine /internal/reload-config: same invalidation for single-process / system-worker-down fallback. Also adds docs/dtype-auto-selection.md (planned design: model-native dtype default read from the checkpoint, precision as explicit override, FA2 fp32 guard) and ignores build artifacts (venv_build.log, .build.pid). Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Audit of the nvidia text loader (cuda.py) vs the shared hf_loading.py found real config drift — the same model config behaved differently on the text path: 1. precision IGNORED — cuda.py hardcoded torch_dtype=float16. Now the manager passes the per-model and cuda.py resolves it via resolve_dtype() (default fp16 on CUDA / fp32 on CPU when unset, preserving current behaviour). 2. 4-bit compute dtype hardcoded float16 → now follows the resolved precision (_make_bnb_config takes a compute_dtype). 3. key was ignored — the manager only read . Now it honours flash_attn / flash_attention (per-model) and the global global_args.flash_attn (offload.flash_attention). 4. offload_buffers was only set on the disk-spill path → now also on the GPU+CPU device_map ladder (with offload_folder), so CPU offload doesn't pin activation buffers on the GPU and OOM the forward pass. 5. global max_ram_gb now clamps the CPU offload budget (central _get_gpu_memory_map_with_limit + the disk fallback), matching hf_loading. Diffusers-only items (component_quantization, GGUF components, sdcpp flash flags) are correctly N/A to the AutoModelForCausalLM text path. Bump version to 0.1.9. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Loading an HF model (e.g. Qwen3.5-9B, 4-bit) failed with 'Permission denied: ./offload'. Cause: the model's per-model offload_dir was the relative './offload' (a stale auto-saved default), which _cfg_or_global lets win over the global config; './offload' resolves to the CWD = the READ-ONLY /opt/coderai/app tree in the image. The config WAS respected — a relative offload path is just meaningless where the CWD isn't writable. * hf_loading.resolve_offload_dir(): an absolute offload_dir is respected as-is; a relative/empty one INHERITS the configured GLOBAL offload directory (global_args.offload_dir) when absolute, then CODERAI_OFFLOAD_DIR, then the user cache — never the CWD. Applied in the manager (both load sites, always passed), hf_loading, and defensively in the cuda backend. * main.py + entrypoint: a container-writable CODERAI_OFFLOAD_DIR (=/cache/offload, created by the entrypoint) is used when the GLOBAL config is still the bare './offload' default; explicit config wins. * run_oci.sh: forward HF_TOKEN / HUGGING_FACE_HUB_TOKEN from the host env so the engines authenticate to the HF Hub (the 'unauthenticated requests' warning) for higher rate limits + gated models. HF_HOME/cache dir was already honoured (main.py from config.models.hf_cache_dir). Bump version to 0.1.8. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The point of the keepalive is to hold the client connection open while the engine is stuck loading — so 'silent' must NOT fall back to the dead legacy path (which sends nothing until the engine's first byte and lets short-timeout clients/proxies disconnect). Now ALL streaming inference goes through the front keepalive path: * silent — SSE comment lines (': …') only: keeps the socket alive, emits no chunk/content/status (invisible to event parsers). * invisible — empty-content chunk + x_queue_info (default). * visible — visible status text. * thinking — reasoning channel (when mode != silent). The front stays responsive even when the engine is GIL-blocked, so it can drive these regardless of engine state. Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
Clients hitting the public API directly (e.g. township via nginx) disconnected during long waits: the direct proxy() path acquired the front queue slot silently and awaited the engine's first byte (model load) with no output. The broker path already kept alive; the direct path now does too. For a STREAMING inference request, commit to a 200 text/event-stream up front and emit keepalive while acquiring the queue slot and during the engine's model load / not-ready retries, then relay the real stream (token-counting for the Tasks page), ending cleanly if the engine dies mid-flight. Configurable mode (models.wait_status_mode, global default 'invisible'; per-model override via the models.json entry): * invisible — empty-content SSE chunk + x_queue_info (holds the connection; no content pollution) * visible — short visible status text (appears in the content) * silent — nothing (legacy path) When thinking is enabled the keepalive goes on the reasoning channel instead (no pollution), unless mode is silent. Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
manager (record_vram_delta): record the FULL footprint after a load — including offloaded loads (no longer skipped) — so the next load skips the OOM→move-to-RAM retry dance: * measured_vram_gb — GPU VRAM delta + reserve (GPU portion; eviction uses it) * measured_ram_gb — host-RAM delta (the CPU-offloaded layers) [new] * measured_n_gpu_layers — the layer split llama.cpp settled on [new] total need = vram + ram. All gated by force_vram_update (else only when used_vram_gb unset). New _learned_n_gpu_layers(): when n_gpu_layers is auto/-1, reuse the learned split so a reload jumps straight to the config that fit. Both load sites snapshot host RAM and read the backend's settled n_gpu_layers. frontproxy: * graceful streaming proxy: when an engine dies mid-stream (e.g. a CUDA ggml_abort SIGABRT on a VRAM-OOM decode), end the relayed stream cleanly instead of throwing an unhandled ASGI exception (noisy traceback). * task actions (cancel/interrupt/pause/resume/restart/DELETE) now route to the engine that OWNS the task — fan out to every live engine (each validates the same session cookie), first success wins — fixing 'Task not found' for tasks on a non-primary engine; front-only synthetic ids handled locally. Bump version to 0.1.7. -
Stefy Lanza (nextime / spora ) authored
* coderai-docker (run_oci.sh): new --tool-arg TOOL VAL (one token) and --tool-args TOOL STR (whitespace string), repeatable, for video-editor | videogen | township | parler. They accumulate into CODERAI_<TOOL>_ARGS env vars that supervisord appends to each tool's command line. Documented in --help (plus the --map dir-or-file note and bring-your-config examples), and shown on the startup banner (tool-args:). * entrypoint: default+export the four CODERAI_*_ARGS (empty) so supervisord's %(ENV_...)s never fails on an undefined var; pre-create /cache/video_editor/sessions. * supervisord: append %(ENV_CODERAI_*_ARGS)s to all four tool launchers; the video editor now runs with --session --session-dir /cache/video_editor/sessions so editor-state recovery is ON by default and persists on the cache volume. * video_editor.py: new --session-dir to place session state/assets at a persistent/mapped path (defaults to ~/.cache/... when omitted). Bump version to 0.1.6. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
coderai always runs as front + engines, so drop the single-process special-casing in execute_api_request: always hand the sub-request to the front (the single API), which routes it to the engine that owns the target model. Critically, remove the on-error fallback to in-process dispatch — that would silently run the model on the WRONG engine and hide a routing failure. If the front can't be located/reached it now returns a 502 the caller surfaces. Drops the now-unused os import. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
* pipelines.py: /v1/pipelines/* (image-to-video, story, video-dub, audio-dub) chained their steps by calling the image/video/text/TTS handlers DIRECTLY in-process, forcing every modality onto whichever engine received the pipeline request. In a multi-engine deployment each model may live on a different engine. Route every sub-step through the front via execute_api_request (same fix as characters/environments) so each lands on the engine that owns that model; in-process fallback for single-process mode. Drops the unused TestClient import. * supervisord.conf: the bundled tools (video_editor, videogen, township) now use --base-url http://127.0.0.1:8776 — the PUBLIC nginx port — so they talk to coderai exactly like any external client (nginx already proxies the full API with 4G bodies / 1h timeouts). The front's own bind stays CODERAI_PORT=18776. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
/v1/characters/generate and /v1/environments/generate produced their reference images by calling execute_internal_request(request.app, '/v1/images/generations') — an IN-PROCESS self-dispatch on whatever engine handled the request. Two bugs: 1. It re-entered the engine's own _InternalAuthMiddleware without the internal token, so in multi-engine mode it 403'd ('engines are reachable only through the front proxy') — image gen failed while plain text worked (text never self-dispatches). Township (a normal API client) hit exactly this. 2. Even with a token it would run the image model on the WRONG engine: the image model may be assigned to / pinned on a different engine than the one serving the characters/environments request. Add execute_api_request(): behind a front (CODERAI_INTERNAL_TOKEN set) it routes the sub-request THROUGH THE FRONT (the single public API) at 127.0.0.1:<front port>, so the front picks the engine that owns the image model and stamps the internal token itself; the caller's Authorization is forwarded so that engine authorises it like any client call. Single-process mode still dispatches in-process (middleware is a no-op with no token). The client never sends internal headers — it just talks to the one API endpoint. Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
Mirror the township config fix for the bundled video editor: * supervisord: pass --config /cache/video_editor/video_editor.config.json so a bind-mounted (or cache-persisted) config auto-loads — previously it only auto-loaded video_editor.config.json from the baked cwd, which isn't mappable. * video_editor.py: a MISSING --config file is no longer fatal (was raise SystemExit) — start with defaults; the web UI Save creates it. * entrypoint: pre-create /cache/video_editor so the config dir exists for auto-load + web-UI Save (persists on the standard /cache volume by default). * run_oci.sh (coderai-docker): --map now accepts a single FILE, not just a dir, so a tool config that lives loose on the host can be linked directly, e.g. --map /host/video_editor.config.json:/cache/video_editor/video_editor.config.json Bump version to 0.1.5. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The bundled township tool (supervisord [program:township]) launched without -c/--config, so a township_config.json bind-mounted via --map into /cache/township_output was never read — the web UI came up with blank API key / league upload credentials / options every restart. * supervisord: pass --config /cache/township_output/township_config.json so a mapped config auto-loads. Map the bare-metal dir with coderai-docker --map /path/to/township_output:/cache/township_output and the saved settings + credentials come up automatically. * gen_township_fighters.py: a MISSING --config file is no longer fatal (fresh install with no mapped dir) — fall back to defaults and start normally; a malformed config still errors. The web UI's Save already writes back to <out-dir>/township_config.json, so first-run saves persist into the mapped dir. Bump version to 0.1.4. Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
The thermal guard was checkpoint-based and ran inside each engine, so it went blind during a single long native call (llama.cpp prefill / image encode): the engine can't reach a between-token checkpoint, and the [thermal][debug] heartbeat stops exactly when the GPU is hottest. Move the AUTHORITATIVE monitor to the front, which stays responsive regardless of what an engine is doing: * Front (engine_supervisor): a thermal thread reads per-card temps via gpu_stats() + CPU temp, maps each card to the owning engine by its CODERAI_ENGINE_GPUS selectors, and drives pause/resume with hysteresis (pause at *_high, resume only back at *_resume). A hot GPU pauses just its engine; a hot CPU pauses all. * Engines stop cooperatively, as before, but triggered remotely: the front POSTs /internal/thermal-pause; thermal.set_external_pause() makes wait_until_safe()/ checkpoint() block at the next safe point (publishing cooldown state so the Tasks page shows it), until /internal/thermal-resume. * Escalation: if a paused engine keeps generating (inflight > 0) — stuck in a native call it can't interrupt — for stop_escalate_checks (default 3) consecutive checks, the front SIGSTOPs the engine's process group; SIGCONT on cooldown. Both signals target the session group so children freeze too. * stop_all()/restart_engine() SIGCONT a frozen engine first (a stopped process ignores SIGTERM until continued); _spawn() resets the thermal flags. * Config: thermal.supervisor_enabled (default on), thermal.stop_escalate_checks. * UI: per-engine temp + pause/frozen state in engines_list and the Tasks cooldown banner (covers a SIGSTOPped engine that can't report its own cooldown). Also: exclude coderai-runtime/*-runtime from the Docker build context (.dockerignore) — a root-owned runtime temp file was breaking the image build. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
record_vram_delta() measured free VRAM via _free_vram_snapshot(), which only read the default CUDA device (device 0). A gpu_split model (e.g. gemma on two cards) lands most weights on the other card, so the device-0 delta came out ~0, tripped the 'delta <= 0' guard, and never persisted measured_vram_gb. The estimate stayed 0 and every load OOM-retried from scratch even with force_vram_update set. Sum free VRAM across every visible CUDA device (scoped per engine via CUDA_VISIBLE_DEVICES), plus amdgpu sysfs when cross-pooling is on or no CUDA device is visible -- mirroring _get_free_vram_gb(). Before/after snapshots are now symmetric, so the delta captures the full multi-card footprint and the real value is persisted; subsequent loads size correctly and stop OOM-looping. Bump version to 0.1.3. Co-Authored-By:
-
- 23 Jun, 2026 10 commits
-
-
Stefy Lanza (nextime / spora ) authored
The loop-guard change added a module-level _GEMMA_CALL_RE = _re.compile(...) at the top of the file, but 'import re as _re' sat far below (line 2266), so import crashed with NameError: name '_re' is not defined and every engine failed to start. Move the import up with the other stdlib imports (removing the late duplicate). Verified codai.api.text now imports cleanly. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
`docker load -i` shows no useful progress and the installer captured its output into a variable, so a ~12G load looked like a multi-minute hang. Stream the tarball through a meter into `docker load` instead — the file read tracks load progress closely. Prefer `pv` (bar + ETA), fall back to GNU `dd status=progress` (bytes + throughput), else plain load with a tip to install pv. The meter draws on stderr so the "Loaded image:" line is still captured. Pre-authenticate sudo first so its prompt doesn't collide with the bar. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The streaming content gate only ran when the request carried a tools array. But the AISBF relay can drop tools (as it does max_tokens), so a gemma model still emitting <|tool_call>call:NAME{…} from its system prompt had that markup streamed straight to the client as content — the spill that poisoned history and fed the tool-call loop. _gate_tool_content now: * ALWAYS withholds the unambiguous native special-token markers (<|tool_call>, <|tool_response>, DeepSeek DSML) and their cross-chunk partials, even with no tools declared; * gates the ambiguous <tool>…</tool> XML form only when tools are present; * gates the gemma-4 call:NAME{…} form by the per-model gemma_tool_parser mode (off: never; full: always; restricted: only declared tool names), so legit call:foo{…} prose/code is streamed instead of withheld+dropped. The gate is now invoked for every stream (not just tools-present), and the final flush likewise, with the mode/tool context resolved once per request. Bump version to 0.1.2. Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
The gemma-4 native call:NAME{…} heuristic could eat legitimate text: the content stripper removed ANY call:/response:{…} span with no tool-name restriction, so a coding reply containing e.g. call:foo{…} in prose or a snippet got silently deleted; and with no tools declared the parser matched any call:word{. Add a 3-way mode, resolved per-model (models.json "gemma_tool_parser") over the global models.gemma_tool_parser (default "restricted"): * full - parse & strip every call:/response: span (old behavior) * restricted - only when NAME is a declared tool; legit call:foo{…} prose/code is preserved, real declared-tool calls still parse & strip * off - disable the gemma heuristic entirely (bigger models that emit standard structured tool calls) resolve_gemma_tool_mode() + a centralized strip_gemma_native() now gate both the parse (GemmaParser, via the dispatcher) and the strip (ModelParserAdapter and the streaming ToolCallParser, which now stashes the declared tool names from extract_tool_calls so 'restricted' works on the streaming finalizer too). Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
A model can get stuck re-issuing the same tool call with the same arguments when each attempt fails (e.g. memory replace old_text=X → "No entry matched" on the wrong memory tier) — or when the call spills as un-parsed call:NAME{…} markup into assistant content. Each brokered request carries the full history, so coderai can see the repetition the agent's own tool-runner didn't break. _detect_tool_loop scans the recent history for a (tool, arguments) signature repeated >= threshold times where those attempts failed (per the tool result) or spilled as markup, and injects one system reminder before generation telling the model to stop repeating the call. Covers both structured tool_calls and spilled gemma-style markup. Configurable on every model: per-model models.json tool_loop_guard / tool_loop_repeats override the global models.tool_loop_* (default on, repeats=3); repeats<=0 disables. Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
ChatMessage.convert_content_array_to_string ran at parse time (mode=before) and unconditionally joined any multipart content list into a string, replacing each image_url part with the literal text "[image_url content]". That destroyed the image before text.py's vision pipeline (_vision_ok / _normalize_vision_content, which the backend's mmproj/MTMDChatHandler consume) ever saw it — so a vision model (e.g. lisa = Gemma-4 + mmproj) received only text and answered "no image attached", and agents looped retrying. text.py's end-to-end vision handling was effectively dead code because content was always pre-stringified. Now flatten only TEXT-ONLY multipart arrays (the KiloCode case); preserve the list whenever it carries an image_url (or any non-text part) so the multimodal backend receives the image. Non-vision models still degrade to the placeholder downstream (text.py:1439), and text-only/plain-string paths are unchanged. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The broker streaming relay counted tokens into engine.active["step"] but never set ["rate"], so _merge_engine_tasks (which overlays both onto the running task) showed token progression with a frozen 0 speed. Compute tok/s from the first streamed token (so the model-load/queue wait doesn't drag the average down) and publish it as m["rate"], mirroring the engine-side text path. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Ships the broker load-status / max_tokens fixes, the run-as-invoking-user default (+ --root opt-out), nginx console silencing, and the expanded installer help. --versioned builds now tag coderai:full_all_0.1.1. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Three operational fixes for the distributable image: run_oci.sh: default the container to the invoking user (uid:gid, SUDO_UID-aware) instead of the image's root default, with a new --root opt-out. The image has no USER directive and supervisord sets no user=, so a run without --user created root-owned dirs (logs, coderai-tmp, hf cache) in the bind-mounted /cache; a later --user run then couldn't write them ("cannot write /cache/logs ... logging to stdout only"). Running as the user by default makes a fresh install safe regardless of run order. --root restores the old behaviour (and throwaway config copy) for shared root-managed data dirs. nginx.conf: error_log crit (was info) + access_log off. nginx's [notice] startup/worker lines and per-request access logs were piped to the container console by supervisord and buried coderai's own output on attach / docker logs. Real failures (crit/alert/emerg) still surface. install.sh: print an extensive post-install guide (quick start, --local + --map, where data lives, the run-as-you default, file logging, docker logs/stop, --help). Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
The engine-side "Waiting for model reply..." content injection ran inside the GIL-blocked engine handler during model load, so it never streamed live -- it only got concatenated into the final reply and polluted the output. Removed it. Load-status now lives in the responsive front-proxy/broker layer: - SSE: broker_execute_stream yields a NON-content chunk (empty delta.content + x_queue_info status) during each not-ready wait (engine down / starting / model loading), so a watching client sees progress without reply pollution. - out-of-band: the existing broker `pending` keepalives are kept. Both are gated by a single flag, resolved by _load_status_enabled(model): - global models.load_status_updates (default True) - per-model "load_status_updates" override in models.json (unset = inherit) The same gate drives the SSE side and, via client.status_gate, the out-of-band `pending` keepalives. Unknown model defaults to on so the relay deadline stays protected. Co-Authored-By:
-
- 22 Jun, 2026 14 commits
-
-
Stefy Lanza (nextime / spora ) authored
Newer OpenAI clients send max_completion_tokens instead of the deprecated max_tokens. Our request model only read max_tokens, so such a request arrived with max_tokens=None and fell back to the default — ignoring the client's limit. Add a max_completion_tokens field and a model_validator that maps it onto max_tokens when max_tokens is omitted (max_tokens wins when both are present). Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
A remote client (Hermes) reportedly sends max_tokens=512, but requests arrive at the engine with no max_tokens (defaulting to 2048). Our broker client preserves whatever payload it's handed, so log the inbound max_tokens / max_completion_tokens and the body keys when a brokered request arrives — this shows definitively whether the value survives the aisbf relay into our payload or is dropped upstream (remote), so we fix the right layer. Diagnostic only. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Two fixes from live debugging: 1. broker_execute_stream: remove the "the model is thinking..." keepalive chunk. It leaked into the assistant reply (saved turns began with the placeholder) and isn't wanted — we just wait for the real response. The engine waits ~5min for a load and the broker sends protocol-level `pending` keepalives, so liveness holds without injecting content. Reverts the streaming body to the simple retry-on-connection/transient-5xx form (no queue/keepalive). 2. _clamp_max_tokens: always resolve a concrete max_tokens, never None. The main stream/generate paths pass request.max_tokens bare; when the client omits it the GGUF backend's `max_tokens or 512` default truncated replies at ~512 tokens mid-sentence. Now: model-level cap is authority (client honored only if smaller); no cap configured -> keep client value, else default 2048. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Add a model-level max_tokens authority: per-model models.json "max_tokens" wins, else a new node-wide models.max_tokens (default None = no cap). A client's requested max_tokens is honored only when it is SMALLER than the model-level cap; a larger or absent request is clamped to the model-level value. Applied before auto-compaction so the window-fit trim works off the clamped value. When no cap is configured, behavior is unchanged (client value, or the 2048 fallback). Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Even after the readiness retries, a brokered stream could still reach the client empty: when the model falls back to CPU offload after a VRAM OOM, the load + first prefill can take tens of seconds during which the engine sends no SSE bytes, so the relay/client times out and shows an empty/errored reply. Keep the stream alive end-to-end: await the engine response and read its stream via a background task, pulling with a timeout. On any gap (long load wait, or slow first token) emit a visible "the model is thinking..." chunk once - a reasoning_content delta when thinking is enabled (enable_thinking/thinking, or a known reasoning model, unless reasoning_effort==none), otherwise a plain content message - then cheap SSE-comment pings to hold the connection without further polluting the reply. Fast responses see no gap and thus no keepalive. Slowness is fine; an empty reply is not. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The first pass only retried when pick_engine returned no engine or the engine returned a non-200. But the real empty-reply case is a CONNECTION failure: during the engine startup/restart window the engine process isn't accepting yet, so self._long.send raises "All connection attempts failed" — caught by the outer except and relayed as a single empty SSE chunk (chunks=1) over the broker. Move the send inside the retry loop on both broker paths (streaming + buffered): retry on a connection exception and on not-ready statuses (425/429/500/502/503/504), attempt-bound (6 x 5s), and only surface the error once exhausted. A 4xx is a real client error and is still relayed as-is. Confirmed against debug.log: all failures were in the engine-startup window; requests after the engine came up streamed fine. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Brokered inference (both the streaming and buffered front executors) hard-failed in the window where no engine/model was ready yet — e.g. engine still warming up, or an OOM-triggered evict+reload in progress. Over the broker that not-ready reply lands at the client as a single empty SSE chunk (the "empty reply / looks like a dead session" symptom), even though a direct API request to the same engine just waits (the engine's own /v1/chat/completions handler retries ~5min for a load). Make the broker paths behave like the direct path: when pick_engine returns no ready engine (or the engine returns a non-200 not-ready status), wait and retry, attempt-bound (6 × 5s), and only surface an error once retries are exhausted. A non-200 means no tokens were generated, so re-sending the body is safe (no duplicate output). Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Root cause (from [timing] logs): chat "build" = 39.6s while prefill+first-token = 0.31s. The ~40s is create_chat_completion rendering the chat template via llama-cpp-python's ImmutableSandboxedEnvironment, whose per-operation security checks make a heavy template (gemma-4's recursive / O(n^2) macros over a big Kilocode conversation) take tens of seconds — every request. Fix (generic for ALL gguf/llama.cpp models): render the prompt ourselves with a plain jinja2.Environment compiled ONCE from the model's tokenizer.chat_template (same trim/lstrip, IgnoreGenerationTags + loopcontrols extensions, tojson filter, same render vars), then tokenize with add_bos=False + special=True (matching llama-cpp's added_special handling — no double BOS) and generate via create_completion. Output is byte-identical to the sandboxed render (verified), but without the sandbox overhead. Plus a small per-message render cache and finer timing ([timing] chat render / tokenize / create_completion setup / first-token). Falls back to create_chat_completion if the model has no chat_template or rendering fails. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The front's broker path buffered the entire SSE response and sent one envelope, so broker clients (lisa, and any OpenAI client via AISBF) got the whole reply at once. Now the front streams it: - streaming.py: stream_chunk_envelope uses event="chunk" with payload.chunk (the shape the AISBF relay consumes), finalize_stream uses event="done". - dispatcher.py: extracted resolve_broker_request() (op routing + body decode + validation) shared by the buffered dispatch and the new streaming path, so they can't diverge; added BrokerDispatchError. - app.py: broker_execute_stream() proxies to the engine with stream=True and yields each SSE chunk (sharing the per-model queue + in-flight tracking); start_broker wires client.stream_dispatcher. - client.py: handle_message, for stream=true requests, relays each yielded chunk as a `chunk` envelope and ends with a `done` (instead of the single buffered reply). Requires the matching AISBF change (send_request returns early for streaming and the relay drains the chunk queue) — committed in the aisbf repo. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Direct Kilocode requests showed ~40s before the first token even though prefill was a cache hit (14551 prefix-match, 1 token to eval) — so the delay is before decode. Added timing in vulkan generate_chat_stream: logs "[timing] chat build (template+tokenize)" around create_chat_completion (eager chat-template render + tokenize) and "[timing] chat lock+prefill+first-token" around the first streamed token. One request now reveals whether the 40s is the gemma chat-template Jinja render (huge template + many tool schemas, done in Python) or lock/prefill. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Capability detection is already name-based (microsecond-fast), so caching it wouldn't help — the real cost is huggingface_hub.scan_cache_dir() + the per-file size sums walking the whole cache. So instead of re-walking on a TTL, the cached-models/cache-stats layer now computes a cheap change-signature (the names + mtimes of the top-level model dirs/files — a handful of stat() calls) and only does the expensive scan when that signature changes. Unchanged caches return the last result instantly and never re-walk; a change refreshes in the background. A long 600s backstop TTL still re-scans occasionally. Mutating ops reset the signature. Net: the models page is instant on every load and the background refresh only runs when models are actually added/removed/downloaded. Verified: unchanged signature → 0 rescans; changed signature → one background rescan. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The models page was slow because the work itself is slow, not because of routing: cached-models walks the HF/GGUF cache doing per-model capability detection, and cache-stats walks every file summing sizes — many seconds each on a large cache. - routes.py: cached-models + cache-stats are now stale-while-revalidate (TTL 25s). First call computes; afterwards they return the last result instantly and refresh in a background thread. Mutating ops (clear-cache, delete cached model, free-disk) force the next refresh. The coderai-system worker pre-warms both scans on startup, so even the first page load is fast. - gpu_detect.gpu_stats(): short 1.5s TTL cache — the Tasks page polls it ~every second and nvidia-smi is slow on a saturated GPU, so collapse repeats. Verified: _cached returns stale instantly and refreshes in the background (2 underlying calls across first+expire); all modules compile. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The front is the authority for config READS (settings GET, status, queue capacity). A settings SAVE stays on the engine because it applies live runtime changes (thermal, RAM monitor, broker runtime) that only exist there and then persists config.json. The front now detects the config.json mtime change and re-reads it into self.config IN PLACE (so the supervisor/admin_data references see it), keeping everything it serves current without a restart. Triggered from status() and the settings GET handler (via an on_config_read callback) and lazily in the capacity helper. This is the pragmatic Phase 2: front owns/serves config; engine applies+persists; front stays consistent. (Fully relocating the 362-line apply logic to the front isn't sound — the live application must run on the engine.) Verified: front reflects an external config.json change (queue_max 6→11) with no restart. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The archive endpoints (list/get/delete/file/settings) are served by the front (codai/frontproxy/admin_data.py, from disk via archive_manager), so the engine's copies in admin/routes.py were dead. Removed. Note: the cache/download/HF handlers in routes.py are NOT removed — the coderai-system worker serves them through the same admin_router, so they're only unrouted on the engine, not dead code. Fully removing them from the engine would require splitting routes.py into engine vs system routers (future cleanup). Co-Authored-By:
-