98 Commits

Author	SHA1	Message	Date
rob thijssen	c8bcaabc38	fix(neuron): render HF chat templates via minijinja pycompat ... The Qwen3.6 chat_template.jinja (now loaded after the precedence fix) failed to render in minijinja: it uses Python str methods (content.startswith/endswith/split/rstrip/lstrip) and the raise_exception global that HF transformers patches into its Jinja env but minijinja doesn't provide. The render error tripped the text-only fallback, so image requests still produced zero <|image_pad|> tokens. Wire the standard bridge into render_chat_template: - minijinja-contrib `pycompat::unknown_method_callback` supplies the Python string/list/dict methods; - a `raise_exception` global maps to a render error (so malformed inputs — e.g. an image in a system message — surface cleanly). Add the real Qwen3.6-27B chat_template.jinja (verbatim from beast's HF cache) as a test fixture and assert it renders one <|image_pad|> for a text+image turn — the end-to-end check that would have caught this before deploy. Refs #16 / TP-vision. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-04 16:32:23 +03:00
rob thijssen	7ad56c6a86	fix(neuron): load chat_template.jinja (transformers precedence) ... The chat-template loader only read the `chat_template` field from tokenizer_config.json. Qwen3.6-27B ships its vision-aware template *only* in a standalone `chat_template.jinja` (and has no tokenizer_config.json at all), so the loader returned None and image requests fell back to the text-only format_qwen3_prompt — rendering zero `<|image_pad|>` tokens and tripping "expand_image_pad_tokens: prompt has 0 image_token_id occurrences". load_chat_template_alongside now follows HF transformers precedence: standalone chat_template.jinja → chat_template.json → the chat_template field in tokenizer_config.json. Tests cover the precedence, the text-only fallback, and that an OpenAI image_url content part renders `<|image_pad|>` through the real template condition (`'image_url' in item`). Refs #16 / TP-vision. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-04 16:25:30 +03:00
rob thijssen	1b0e36c119	fix(neuron): cover TpForwardLogitsWithImages in drain_poisoned match ... The CUDA type-check caught a non-exhaustive match: drain_poisoned() must reply an error to every Job variant's reply channel, including the new cuda-gated TpForwardLogitsWithImages. The non-cuda build couldn't see it — the variant is #[cfg(feature = "cuda")], so the match is exhaustive without it on CPU. Refs TP-vision plan Stage 2. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-04 15:26:46 +03:00
rob thijssen	ed2d09864e	feat(neuron): TP-vision Stage 3 — wire TP chat + stream vision prefill ... End-to-end TP-vision: an image request to a TP-loaded Qwen3.6-27B now conditions on the image across both ranks. - TpLoadedModel carries has_vision / image_token_id / lm_tokens_per_image, populated at load via the shared VisionMeta::from_config_path (same config.json the shards loaded from; Stage 1 materialises the replicated tower on every rank). - LoadedHandle::capabilities() now advertises "vision" for TP loads with a tower (cortex-gateway already unions this into /v1/models via C3). - The TP rejection guards (chat_completion_tp + inference_tp_stream) are now conditional on !has_vision — text-only TP models still 400 cleanly, vision-capable ones fall through. - chat_completion_tp_inner and the streaming orchestration task detect images (request_has_images), expand <|image_pad|> to the per-image patch count, and run a single-shot generate_step_with_images prefill (every rank encodes + splices its replicated tower) before the unchanged decode loop. Text requests keep chunked_prefill_tp. - extract_image_data_uris ships the source data URIs to every rank for identical per-rank preprocessing. prompt_tokens now reflects the patch expansion, so usage accounting and KV offsets match the single-GPU baseline. TP entry points are cuda-gated (validated by CI's CUDA type-check); capabilities() + extract_image_data_uris + VisionMeta reuse compile on the non-cuda build. Full workspace test green. Refs TP-vision plan Stage 3. Implements #12. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-04 15:14:44 +03:00
rob thijssen	4994b94c84	feat(neuron): TP-vision Stage 2 — per-rank image RPC + worker plumbing ... Carry image content through the TP forward path so every rank encodes and splices locally (replicated tower, no embedding broadcast). - rpc.rs: new WorkerRequest::GenerateStepWithImages carrying the source image data URIs + image_token_id for the single-shot vision prefill; worker still replies GenerateStepOk. Round-trip test added. - tp_qwen3_5.rs: TpQwen3_5ForCausalLM::forward_with_images — encode each preprocessed image through the rank's replicated tower, cat, splice, forward. Shared by leader and worker so every rank runs identical work. - tp/mod.rs: TpLeaderModel::forward_with_images and WorkerPool::generate_step_with_images (mirrors generate_step: fan out GenerateStepWithImages to subprocess ranks, run the leader's image forward on its device worker thread, drain, combine). - worker.rs: WorkerModel::forward_with_images + handle_generate_step_with_images — each subprocess rank preprocesses the same data URIs via the shared deterministic preprocess_data_uri, encodes, splices, forwards. - device_worker: Job::TpForwardLogitsWithImages + tp_forward_logits_with_images dispatch handler + DeviceWorkerHandle::tp_forward_logits_with_images. Determinism: every rank runs the same preprocess on the same source URIs through the same replicated tower, so the spliced hidden state matches across ranks — preserving the replicated-hidden-state invariant the row-parallel AllReduce relies on, with no NCCL broadcast. No caller yet — Stage 3 wires the TP chat/stream entry points to invoke generate_step_with_images for image prefill. cuda-gated plumbing covered by CI's CUDA type-check; rpc/route/forward_with_images compile on the non-cuda build. Refs TP-vision plan Stage 2. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-04 15:08:08 +03:00
rob thijssen	9a24b05866	feat(neuron): TP-vision Stage 1 — replicated vision tower on the TP model ... Load the full, unsharded model.visual.* vision tower on every TP rank (leader + each subprocess worker mmaps the same local safetensors) when config.vision_config is present. VisionTower::load already takes a ShardedVarBuilder whose plain .get() returns the full replicated tensor, so the tower loads identically regardless of world_size — no sharding, no NCCL broadcast. - TpQwen3_5ForCausalLM gains vision: Option<VisionTower> + image_token_id, plus has_vision/image_token_id/encode_image/forward_with_vision, mirroring the single-GPU Qwen3_5ForCausalLM wrapper. - TpQwen3_5Model::forward_with_vision mirrors the single-GPU forward_inner splice: embed locally, replace rows at image_token_id positions, run the sharded decoder stack. Because every rank encodes the same pixels through its replicated tower, the spliced input embeddings are identical across ranks — preserving the TP replicated-hidden-state invariant the row-parallel AllReduce relies on. - splice_runs is now pub(crate) and shared with the TP model. No caller yet — Stage 2 wires the RPC/worker path that invokes encode_image + forward_with_vision per rank. Most of this compiles on the non-cuda build (only the cuda load variant's tower line is gated); CI's CUDA type-check covers the rest. Refs TP-vision plan Stage 1. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-04 15:00:05 +03:00
rob thijssen	f8c0da0ebf	fix(neuron): TP-vision Stage 0 — reject image requests on the TP path ... The TP inference path has no vision tower, and the TP dispatch in chat_completion / inference_stream returns before the VisionUnsupported guard runs — so an image request to a TP-loaded model (e.g. beast's tp=2 Qwen3.6-27B) was silently dropped and answered from text alone, the exact issue-#3 confident-hallucination pattern Stage C killed for single-GPU. Add the request_has_images → VisionUnsupported guard to both chat_completion_tp and inference_tp_stream, before prefill / before the SSE stream opens, so beast returns a clean 400 vision_unsupported. The guard is unconditional for now (TP has no tower); Stage 3 makes it conditional on the TP model's has_vision once real TP-vision lands. Detection is covered by the existing request_has_images unit test; the guard itself is cuda-gated (validated by CI's CUDA type-check). Refs TP-vision plan Stage 0. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-04 14:53:56 +03:00
rob thijssen	dd592d918d	test(neuron): C2 — guard Responses→chat image translation contract ... The Responses request translator already emits the chat `image_url` Parts array Stage B5's vision path consumes, and the non-streaming (`chat_completion`) and streaming (`responses_stream` → `inference_stream`, Stage C1) Responses paths both route image content to the vision-aware prefill — so vision works end-to-end through `/v1/responses` with no translator change required. Add a multi-image test asserting order preservation and that the `detail` hint is tolerated (and dropped, since chat image_url has no analogue), locking the translator's output to the exact `image_url.url` shape `extract_images_from_request` walks. Closes part of #16 (Stage C2). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-04 13:57:43 +03:00
rob thijssen	766c20ba47	feat(neuron): C1 — streaming SSE chat completion with vision ... The streaming worker path now splices image embeddings on prefill, closing the silent text-only degrade for `stream=true` image requests. `inference_stream` gains the same vision-routing block as the non-streaming `chat_completion`: detect `image_url` content, reject it against text-only models with `VisionUnsupported` (before any SSE frame is sent), preprocess each image and expand its `<|image_pad|>` sentinel to the per-image patch count, then carry the payload through dispatch. Rather than duplicate the 75-line `route_token!` reasoning/tool-call state machine into a sibling streamer, `stream_inference_via_worker` takes an `Option<(Vec<ImageInput>, u32)>`: when `Some`, prefill is a single-shot `forward_logits_with_images` splice; when `None`, the original chunked text-only prefill. Image embeddings are prefill-only, so every decode step stays on the plain `forward_logits` path and the shared decode loop is untouched. This keeps exactly one copy of the tool-call/reasoning logic to maintain. The Responses API streaming path (`responses_stream`) inherits vision for free since it drives the same `inference_stream`. Unit test covers `request_has_images` (the shared routing gate); the real-weights SSE smoke is the manual curl on beast (cuda-integration). Closes part of #16 (Stage C1). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>	2026-06-04 13:57:02 +03:00
rob thijssen	577781de8d	fix(neuron): derive Clone on ImageInput for the CUDA vision dispatch ... CUDA type-check in CI failed on commit 24968e9 with E0308: error[E0308]: mismatched types --> crates/neuron/src/harness/candle.rs:1707:33 1707 | images.clone(), | ^^^^^^^^^^^^^^ expected `Vec<ImageInput>`, found `&Vec<ImageInput>` In Stage B5 the cuda branch of `chat_completion` matches `&vision_route` to keep the `vision_route: Option<...>` alive for both arms, which makes `images` bind as `&Vec<ImageInput>`. The subsequent `images.clone()` call doesn't deep-clone because `ImageInput` doesn't derive `Clone` — rustc falls back to cloning the `&Vec` reference, which has the wrong type for the worker job. The CPU build (non-cuda) compiled fine because that branch is behind `#[cfg(feature = "cuda")]`; the cuda-check job is what catches the regression. Fix: derive `Clone` on `ImageInput`. The clone cost is one pixel-buffer memcpy per image (~2.4 MiB at fixed 448×448), which is fine on the chat-completion hot path — vision requests are rare per second relative to text-only decode. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-06-02 15:51:57 +03:00
rob thijssen	24968e9233	feat(neuron): Stage B — end-to-end text+image chat for Qwen3.6 ... Stage B of the vision plan (doc/vision-qwen3_6-spec.md). Wires the vision tower from Stage A through to a complete non-streaming chat completion: extract images from the request, preprocess, encode on the worker thread, splice embeddings into the LM input at `<|image_pad|>` positions, return coherent text response with `prompt_tokens` reflecting patch tokens. Closes the silent-drop class of failures from issue #3 — vision requests against Qwen3.6 now condition the model on the image instead of producing confident text-only hallucinations. Streaming for vision is Stage C. Deferred items tracked under #12 (TP-vision), #13 (27B production), #14 (dynamic resolution), #15 (numerical validation). What landed: - **B1 — `Qwen3_5Model::forward_with_vision`**: text-only `forward` unchanged; new method takes `(input_ids, offset, image_embeds, image_token_id)`, embeds tokens, locates `image_token_id` positions, splices via the new `splice_runs` helper. MRoPE applies text-positions to image tokens for Stage B (spatial MRoPE is the issue #15 numerical-validation follow-up). 2 unit tests for `splice_runs` covering contiguous + non-contiguous runs. - **B2 — `ModelArch::forward_with_vision` dispatch**: routes Qwen3_5Dense to the new method; other arches return an error. Defence-in-depth — the HTTP layer (B6) already rejects image content for non-vision models. - **B3 — `Job::ForwardLogitsWithImages`**: new worker variant carrying tokens + per-image `(pixels, c, h, w)` payloads. The dispatcher encodes each image (device-resident), concatenates the resulting embeddings, calls `arch.forward_with_vision`, and returns CPU logits. Image embeddings never copy back to CPU — the "tensors don't escape the worker" invariant from the per-device worker refactor still holds. Poisoned-worker drain path handles the new variant. - **B4 — Prompt builder**: - `request_has_images` detects image content cheaply. - `extract_images_from_request(request, profile)` walks `MessageContent::Parts`, decodes data URIs, runs `harness::preprocess::preprocess` per image, returns `Vec<ImageInput>` in request order. - `expand_image_pad_tokens(input_ids, image_token_id, patches_per_image)` walks the tokenized prompt and replaces each `<|image_pad|>` (id 248056 for Qwen3.6) with N copies matching the per-image patch count. 4 unit tests. - `VisionMeta::from_config_path` peeks `config.json` at load time for `image_token_id`, vision_config patch/merge sizes, and derives `lm_tokens_per_image` for the Stage B fixed resolution. - **B5 — `chat_completion` vision routing**: detects image content, validates the loaded model has vision, expands the prompt, and calls a new `run_inference_with_images_via_worker` helper that does single-shot prefill + standard decode loop (KV cache holds the post-splice hidden states from prefill, so decode steps don't re-splice). Stage B skips chunked prefill for vision — at 448×448 fixed resolution the budget stays well under the activation-memory threshold. Long-vision chunking is Stage D follow-up. - **B6 — `InferenceError::VisionUnsupported`**: structured 400 with `code=vision_unsupported, model_id, suggestion` when an image request hits a non-vision model. Closes the agent0 failure mode where vision requests degraded silently. - **B7 — `ModelInfo.capabilities`**: per-model array (`["text"]` vs `["text", "vision"]`) in `/v1/models` and forwarded verbatim by cortex-gateway. Lets clients (litellm, agent0) gate image_url submission on the declared capability set. Optional in the wire format; defaults to empty for older clients. CI gate: cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace (all 28 test groups ok, 124 lib tests). New unit-test counts: +2 splice_runs, +4 expand_image_pad. Manual verification (after RPMs deploy on beast): curl http://hanzalova.internal:31313/v1/chat/completions \ -H 'Content-Type: application/json' \ -d "{\"model\":\"Qwen/Qwen3.6-27B\", \"messages\":[{\"role\":\"user\",\"content\":[ {\"type\":\"text\",\"text\":\"What's in this image?\"}, {\"type\":\"image_url\",\"image_url\":{\"url\":\"data:image/jpeg;base64,...\"}} ]}], \"max_tokens\":120}" | jq Expect prompt_tokens > 196 (text + 196 patch tokens) and a response that references actual image content. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-06-02 15:33:00 +03:00
rob thijssen	7df84fed8f	feat(neuron): Stage A — vision tower load + preprocessor for Qwen3.6 ... Stage A of the vision implementation plan (doc/vision-qwen3_6-spec.md). Builds the vision tower scaffolding that today's silent-drop failure mode (issue #3) needs — the Qwen3.6 ViT loads from `model.visual.*`, runs forward producing post-merger LM-side image embeddings, and routes through the device worker via a new `Job::EncodeImage`. No LM splice yet — that's Stage B. Refs #3 (umbrella). Deferred sub-stages tracked as #12 (TP-vision), #13 (27B production deploy), #14 (dynamic resolution), #15 (numerical validation). What landed: - **A0 — investigation**: pulled config.json, preprocessor_config.json, chat_template.jinja, and safetensors index from beast's local Qwen3.6-27B cache. Documented in doc/vision-qwen3_6-spec.md with exact tensor shapes for every `model.visual.*` weight. Confirms 27-block ViT with `hidden_size=1152`, `patch_size=16`, `spatial_merge_size=2`, `out_hidden_size=5120`. Vision tower lives in 2 of the 15 safetensors shards. - **A1 — deps + scaffolding**: added `image = "0.25"` (default- features off, PNG/JPEG/WebP/BMP/GIF) and `base64 = "0.22"` to crates/neuron/Cargo.toml. Created `harness::preprocess` and `harness::arch::qwen3_5::vision` modules. - **A2 — preprocess.rs**: `decode_data_uri` strips `data:image/...;base64,...` → image bytes → `image::DynamicImage` (rejecting `http(s)://` URLs to avoid SSRF/recursion); `preprocess` resizes to a fixed `PreprocessProfile::qwen3_6()` (448×448), normalises to `[-1, 1]` per the model's mean/std=0.5, emits row-major `(3, H, W)` f32. 9 unit tests covering data URI parse, decode failure paths, grayscale-to-RGB promotion, and the exact-value normalisation contract. - **A3 — vision.rs**: `VisionTower` struct with `patch_embed: Conv2d`, learned `pos_embed: Embedding`, 27 `VisionBlock`s (pre-LN + multi-head self-attention with fused QKV + GELU-tanh MLP + residuals), and `VisionMerger` (LayerNorm → 2×2 spatial concat → linear_fc1 → GELU-tanh → linear_fc2 to LM hidden_size). Includes the Conv3d→Conv2d fold trick documented at the top of the file — the published patch_embed.proj.weight is 5D `(1152, 3, 2, 16, 16)` but candle 0.10 has no Conv3d; for static images we sum-collapse the temporal axis. Video would need real Conv3d. 5 unit tests including the exact `gelu_pytorch_tanh` reference values from PyTorch. - **A4 — wire vision into Qwen3_5ForCausalLM**: extended `Config` with optional `vision_config: Option<VisionConfig>` and `image_token_id`; `Qwen3_5ForCausalLM::new` now loads the vision tower when present, exposes `has_vision()` and `vision()` so the HTTP layer can advertise capability and so the encode path can reach it. - **A5 — device worker `Job::EncodeImage`**: new job variant carrying CPU-side `(C, H, W)` pixels. Dispatch handler reconstructs the tensor on the worker's device, calls `arch.encode_image(image)`, copies the result back to CPU as flat `Vec<f32>`. Keeps the "tensors don't escape the worker" invariant. Poisoned-worker drain path handles the new variant. - **A6 — dispatch round-trip test**: `encode_image_routes_to_dispatch_ and_errors_on_unknown_handle` proves the channel/dispatch wiring works end-to-end via the CPU device worker (errors on unknown ArchHandle, which is the expected behaviour without a loaded model — real-weights validation happens in Stage B when the LM splice path exists). CI gate: cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace (all 28 test groups ok, zero failures). New test counts: +9 in preprocess, +5 in vision, +1 in device_worker. Out of scope (deferred): - LM-side splice of image embeddings at `<|image_pad|>` positions → Stage B. - Streaming SSE for vision-bearing chat completions → Stage C. - Reject `image_url` with HTTP 400 for non-vision models / advertise `capabilities` in /v1/models → Stage C. - TP-vision (#12), 27B production deploy (#13), dynamic resolution (#14), numerical validation (#15). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-06-02 11:40:47 +03:00
rob thijssen	d4e1b05956	feat(neuron,cortex-core): source-aware loader (scheme:org/name) ... Phase 1 of plan-source-aware-loader-preflight. Makes neuron's loader treat `huggingface:org/name` and `helexa:org/name` as first-class distinct sources with per-source endpoint + cache, while staying backwards-compatible with bare `org/name` ids. Zero behavior change for existing operator configs. Motivation: helexa is adding an EU-hosted registry (`registry.helexa.ai`) alongside HF. Both speak HF-compatible wire format, but the bytes, jurisdiction, trust root, and cache namespace are distinct. The loader needs to disambiguate which registry serves a given model id, and to keep their caches from colliding on disk when both happen to host the same `org/name`. What lands: - `cortex-core::source` — new module. `ModelSourceId { scheme, org, name }` with `FromStr` accepting both `scheme:org/name` and bare `org/name`. `Display` round-trips. `repo_path()` emits the `org/name` half for the hf-hub `Api::model(...)` call regardless of which scheme/endpoint we're hitting. Rejects malformed input with typed `ParseError` variants (empty scheme, missing slash, scheme with `/`, name with `:`, etc.). - `neuron::config::CandleHarnessConfig` gains `default_source: Option<String>` and `sources: HashMap<String, SourceConfig>`. `SourceConfig` mirrors what `hf_hub::ApiBuilder` consumes: endpoint URL, optional `auth_env` (env var name read at startup so secrets stay out of TOML), and optional cache_dir. Defaults synthesise a `huggingface` entry pointing at `https://huggingface.co` with the legacy `hf_cache` field as its cache_dir — so existing configs that only set `hf_cache` keep working unchanged. - `CandleHarness::new(bind_url, &CandleHarnessConfig)` replaces `CandleHarness::new(bind_url, hf_cache)`. Resolves every configured source's auth env var and cache dir up front so `hf_api_for(scheme)` is a pure HashMap lookup on the hot load path. Only the `huggingface` scheme gets the legacy `HF_HUB_CACHE`/`HF_HOME` env-var fallback chain; other schemes resolve to whatever the operator typed. - `hf_api()` -> `hf_api_for(scheme)`. Builds an `hf_hub::Api` with the source's endpoint, cache_dir, and auth token. Errors with a useful message naming the configured schemes when an unknown scheme is requested. - `CandleHarness::load_model` parses `spec.model_id` into a `ModelSourceId`, substitutes `default_source` for bare ids, and threads the parsed source through `preflight`, `resolve_files`, `resolve_dense_files`, `load_arch_gguf`, `load_arch_dense`, and `load_tp`. The hf-hub `Api::model()` call now uses `source_id.repo_path()` so registry calls hit the right URL shape regardless of scheme. - `preflight()` signature gains a `&ModelSourceId` parameter (it's the canonical id for log lines and error display); `RepoFetchFailed.model_id` etc. now carry the scheme-qualified form so operator-visible errors echo exactly what was configured. - `neuron.example.toml` documents the new `[harness.candle.sources.*]` table with commented-out examples for `huggingface` (explicit override) and `helexa`. Tests: - 13 new unit tests in `cortex-core::source` covering parse / display round-trip, default-scheme substitution semantics, and every `ParseError` variant. - 6 new unit tests in `neuron::config` covering the `effective_sources` synth (legacy `hf_cache` carry-through, explicit override preservation, helexa-alongside-huggingface) and `effective_default_source` fallback. - 2 new unit tests in `harness::candle::tests` covering multi-scheme `hf_api_for` routing, including the "unknown scheme" error path naming configured schemes. - Preflight integration tests updated to construct `ModelSourceId` and assert against the scheme-qualified error form. CI gate: cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace (all 24 test groups ok, zero failures). Out of scope (Phase 3): - Cortex catalogue `source` field — independent of Phase 1+2, ships when the registry comes online. - `helexa` source endpoint itself — separate project; this PR adds the client-side rails only. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-06-01 13:42:11 +03:00
rob thijssen	61adff347a	feat(neuron): preflight placement check with structured errors ... Phase 2 of plan-source-aware-loader-preflight. Adds a one-RTT placement feasibility check that runs before any device allocation, NCCL handshake, or weight fetch. Replaces today's opaque "fetch config.json … 404" failure mode (when an operator points `tensor_parallel = 2` at a GGUF-only repo) with a structured error that names the failure class and points at the fix. What lands: - `crates/neuron/src/harness/preflight.rs` — new module. Classifies a repo's siblings listing into `SourceFormat` (Gguf | DenseSafetensors | Mixed | Empty), applies the tp/quant feasibility table, returns a `PlacementPlan` on success or a typed `PreflightError` on rejection. `PreflightError` is `serde::Serialize` so the HTTP layer can emit the structured shape verbatim; it's `thiserror::Error` so log lines get a single-line Display when downcasting from anyhow. Includes best-effort Levenshtein-nearest suggestion for malformed quant names (the second sharp edge the HauhauCS scenario surfaced — operator writes `q6k` against filenames containing `Q6_K_P`, and today's matcher just says "no GGUF file matching quant"). - `CandleHarness::load_model` — calls `preflight(...)` first thing after the "already loaded" guard, before any `ensure_device_worker` or `resolve_*`. Failure wraps the typed error in `anyhow::Error` so the existing trait surface is unchanged; the HTTP handler and the startup logger downcast to recover the structured form. - `crates/neuron/src/api.rs::load_model` handler — maps `PreflightError` to 422 Unprocessable Entity with `{"error": {"kind": "...", "model_id": "...", "suggestion": "..." }}`. Other failures keep the existing 400 + free-form `format!("{e:#}")` shape. - `crates/neuron/src/startup.rs::load_default_models` — when the failure is a preflight rejection, log as `reason=<kind> detail=<msg>` instead of the opaque `error=<chain>`, so journalctl on beast will now show `reason=tp_requires_safetensors detail="repo is GGUF-only (8 .gguf files); TP requires dense safetensors..."` instead of `error=fetch config.json from HauhauCS/...: 404 Not Found`. Tests: - 18 unit tests in `harness/preflight.rs` covering classifier, quant matching, Levenshtein, error serialization, and the full feasibility table (gguf+tp rejected, gguf+bad-quant suggests nearest, gguf+good-quant ok, dense+tp ok, empty rejected, mixed prefers safetensors). - 7 integration tests in `tests/preflight.rs` exercising the network path through an axum mock that serves hf-hub-compatible `/api/models/{org}/{name}/revision/main` payloads. Adds `tempfile` as a dev-dependency for per-test cache dirs. Out of scope (deferred to subsequent phases): - Phase 1 (source-aware loader plumbing — `scheme:org/name` parsing, per-scheme `SourceConfig`, cache disambiguation). Preflight runs against the single configured HuggingFace source today; the scheme threading lands cleanly when Phase 1 ships. - Phase 3 (cortex catalogue source field). - GGUF tensor-parallel loading. Preflight rejects this combination with `TpRequiresSafetensors`; the underlying loader gap is the separate `Helexa` curated-registry / heretic-rs conversation. Refs #4-#9 architectural follow-up; no specific issue closed. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-06-01 13:24:30 +03:00
rob thijssen	435fd10902	fix(neuron): macro-ify CUDA single-GPU route_token so DecodeStream type stays inferred ... Prerelease build (run 270) failed on commit cb30383 with: error[E0107]: struct takes 5 generic arguments but 0 generic arguments were supplied --> crates/neuron/src/harness/candle.rs:3554:41 | 3554 | decode_stream: &mut tokenizers::DecodeStream<'_>, | ^^^^^^^^^^^^ The Step-2-era refactor for #6's tool-call extraction added a nested `async fn route_token` inside `stream_inference_via_worker` that named `tokenizers::DecodeStream<'_>` as a parameter type. `DecodeStream` actually has five generic parameters (`'tok, M, N, PT, PP, D`) which makes naming it explicitly painful — the working approach the CPU path uses is a macro, where the body expands inline at the call site and the decoder type stays inferred. This commit replicates the CPU-side macro for the CUDA worker path. Same shape, just with `.await` calls inside (macros tolerate that since they expand inline into the enclosing async context). Control flow uses a labelled-block + `consumer_alive` flag rather than `return` so the macro stays generic over the surrounding return type. The CPU build (default-feature workspace, what `clippy` and `test` jobs exercise) doesn't compile this `#[cfg(feature = "cuda")]` branch, which is why local CI green-lit it. The cuda-check job should catch this category of breakage now that #cb30383+CI-fix landed; this commit just resolves the actual breakage on the prerelease workflow. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-06-01 08:59:56 +03:00
rob thijssen	cb303832bc	feat(neuron): render the model's chat_template with chat_template_kwargs ... Closes #9. Replaces the hardcoded `format_qwen3_prompt` ChatML glue with `minijinja`-driven rendering of the model's own `chat_template` from `tokenizer_config.json`. The request's `chat_template_kwargs` flow into the Jinja context so model-specific levers (Qwen3's `enable_thinking: false`, etc.) actually take effect. ## Implementation - New `harness::chat_template` module with three entry points: - `load_chat_template_alongside(tokenizer_json_path)` — probes `tokenizer_config.json` in the same hf-hub snapshot directory. Supports both the canonical string-form `chat_template` and the array-form some tokenizers ship (multi-template models). - `render_chat_template(template, messages, tools, kwargs)` — renders via `minijinja`. Messages flatten into the `[{role, content}]` shape HF templates iterate, with per-message extras (`tool_calls`, `tool_call_id`) preserved. `tools` and `kwargs` add into the Jinja context so templates that reference them work without us interpreting their shape. - `chat_templates_enabled()` reads `NEURON_USE_CHAT_TEMPLATE` (default true). Falsy values force the fallback path everywhere — a kill switch for emergency rollback without a rebuild. - `LoadedModel.chat_template: Option<String>` and the TP equivalent are populated once at load time. `None` (no tokenizer_config.json, parse error, missing field) routes the fallback path silently; logs go through `tracing::debug`/`warn` per condition. - New `build_prompt_for_request(chat_template, request)` wraps the decision: when both the template is present AND the kill switch is off, render with kwargs from `request.extra` (looks up `chat_template_kwargs` and `tools` lazily). On render error → warn + fallback to `format_qwen3_prompt`. Wired into all four current prompt-build sites (single-GPU stream + non-stream, TP stream + non-stream). ## Dependency `minijinja = "2"` with the `builtins`, `json`, and `serde` features. Pure-Rust Jinja2 implementation, ~80KB compiled. Used internally by HF's `tokenizers-rs` for its own chat templating; the API surface we touch (`Environment::add_template` + `Template::render(serde_value)`) is stable. ## Validation strategy I can't byte-compare the new path's output against `format_qwen3_prompt` for live models without GPU (CI doesn't have one). The fallback path and kill switch are the mitigations — a deploy can flip `NEURON_USE_CHAT_TEMPLATE=false` in the neuron service env if the chat template renders surprisingly on Qwen3-8B in production. The legacy formatter stays the fail-closed default. ## Scope cuts (documented in module header) - Tool-definition lifting from helexa-acp's system-prompt injection into the chat_template's native tools block is deferred. Today the request's `tools` array threads into the Jinja context, but helexa-acp continues to inject Hermes-format tool descriptions into the system prompt for backwards-compat with non-cortex endpoints. ## Tests 9 unit tests in `chat_template`: kill-switch matrix (truthy / falsy / unset), template loading (string form, array form, missing file, unparseable JSON, missing field), rendering (basic conversation threading, kwargs forwarding, message-extras threading for tool_calls). 215 workspace tests pass; clippy + fmt clean across all workspace features (default). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-31 23:43:11 +03:00
rob thijssen	44008358c5	feat(neuron): emit response.in_progress between created and output_item.added ... Refs #7. OpenAI's Responses API spec emits `response.in_progress` between `response.created` and the first output-item event to mark "request validated, model is generating". Some Responses-API clients distinguish loading-spinner vs streaming-spinner UI based on which event arrived last; emitting both keeps the wire shape matched. Carries the same shell as `response.created` (status=in_progress, empty output, no usage yet) — both events are payload-light bookkeeping, distinguished only by the event name. The hosted-tool event families remaining in #7 (web_search_call, code_interpreter_call, file_search_call, image_generation_call) stay deferred until the underlying tools exist in neuron. Updated `full_stream_emits_expected_event_sequence` to assert the new event lands in position 1; downstream indexing shifted by one across the existing test assertions. CI green, fmt + clippy clean. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-31 23:30:34 +03:00
rob thijssen	fc9a8c42a3	feat(neuron): extract <tool_call> blocks to structured tool_calls deltas ... Closes #6. Same model-agnostic seam as #8 but for tool-call markers (`<tool_call>` / `</tool_call>` on Qwen3-Coder, Hermes-format, DeepSeek-Coder, gpt-oss, …). Lets Zed's tool-use feature and any other vanilla OpenAI chat client get structured `tool_calls` deltas out of cortex without having to parse markers themselves. ## Implementation 1. **Tokenizer probe at load time** (`detect_tool_call_token_pair` in `wire::event`) — same shape as the reasoning-marker probe from #8. Both open AND close must resolve to single token ids; non-tool-use models get `None` and pass through unchanged. Stored on `LoadedModel.tool_call_tokens` and the TP analogue. 2. **New `InferenceEvent::ToolCall` variant** — carries `index` (call slot, per-turn counter), generated `id` (`call_<hex>_<idx>`), `name`, and the complete `arguments` JSON string. One event per parsed call. 3. **Token-level state machine** in all three streaming paths (CPU `run_inference_streaming`, CUDA single-GPU `stream_inference_via_worker`, CUDA TP `chat_completion_tp_stream`) layered on top of #8's reasoning routing: - `<tool_call>` token → enter buffering state, clear buffer. - Tokens while buffering → accumulate into `tool_call_buf` via the decoder (so multi-byte UTF-8 still buffers correctly) without emitting anything visible. - `</tool_call>` token → take the buffer, parse with `parse_tool_call_body` (extract `name` + `arguments`), emit a structured `ToolCall` event with a fresh `call_<hex>` id and the parsed fields. - On parse failure → fall back to re-emitting the original `<tool_call>{buf}</tool_call>` block as plain text content so helexa-acp's existing `ToolCallParser` repair passes still have a chance to recover the call. 4. **OpenAI chat projector** emits the OpenAI streaming `tool_calls` delta shape on `InferenceEvent::ToolCall` — `{tool_calls: [{index, id, type:"function", function:{name, arguments}}]}`. One chunk per call slot. 5. **OpenAI Responses projector** drops `ToolCall` events for now (Responses-side function_call event family routing tracked under #7); the chat path is what unblocks Zed's tool use today. ## Acceptance - Vanilla OpenAI chat clients (Zed's tool-use feature, any other OpenAI-compatible tool-call consumer) get structured tool_calls deltas against cortex+neuron without having to parse `<tool_call>` markers in content. - helexa-acp continues to work — when neuron parses cleanly, it consumes the structured deltas through its existing decoder. When the model emits malformed JSON, neuron falls back to text pass-through and helexa-acp's `ToolCallParser` recovers via the same path it always did. - Models without tool-call markers in their tokenizer pass through unchanged. - No hardcoded model knowledge — entirely driven by tokenizer metadata. ## Tests 2 new detection tests in `wire::event` (Qwen3-style marker detection, no-marker case). The streaming paths themselves stay covered by the existing chat-completions integration tests; full end-to-end exercise of the new path requires GPU-loaded models and lives outside the CI test surface. 215 workspace tests pass; clippy + fmt clean across the workspace. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-31 23:26:31 +03:00
rob thijssen	7733eecba5	feat(neuron): strip reasoning from chat completions by default ... Closes #8. Reasoning-capable models (Qwen3, DeepSeek-R1, gpt-oss, Mistral Magistral, …) emit `<think>...</think>` blocks inline in their content stream. The chat-completions wire format has no slot for reasoning, so until this change every consumer either parsed the markers themselves (helexa-acp) or wrote the raw scratchpad content into their UI (Zed's commit-message generator — visible as the leaked reasoning block on every generated commit message against benjy's Qwen3-8B). ## Implementation, model-agnostic by design The neuron side now does token-level routing without any hardcoded model knowledge: 1. **At load time** (`detect_reasoning_token_pair` in `wire::event`), probe the tokenizer's vocabulary for a known reasoning-marker pair: `<think>` / `</think>` (Qwen3, DeepSeek-R1, gpt-oss), `[THINK]` / `[/THINK]` (Mistral Magistral), and a couple of derivatives. Each marker must resolve to a single token id; if both open and close resolve, stash on `LoadedModel.reasoning_tokens` (similarly `TpLoadedModel`). Non-reasoning models get `None` and pass through unchanged. 2. **At inference time**, the three streaming paths (`run_inference_streaming` CPU, `stream_inference_via_worker` CUDA single-GPU, `chat_completion_tp_stream` CUDA TP) now check each sampled token against the pair via the new `handle_reasoning_marker` helper before feeding it to the detokeniser. Open marker → set `in_reasoning = true`, drop the marker. Close marker → unset, drop. Other tokens go through `emit_delta(_blocking)` which now picks `ReasoningDelta` or `TextDelta` based on state. Markers never appear in the streamed output. 3. **In `wire::openai_chat`**, the projector splits into: - `project_chat_stream` (unchanged signature; default behaviour — drops `ReasoningDelta`) - `project_chat_stream_with(rx, …, ChatProjectionConfig)` — when `include_thinking: true` and `reasoning_markers: Some(_)`, re-wraps reasoning content with the literal open/close marker text and emits as content deltas. Preserves the on-the-wire shape that helexa-acp's `ThinkParser` expects. 4. **HTTP handler** reads `x-include-thinking: true` (case- insensitive `1`/`true`/`yes`) from the request headers and threads it into the projection config. cortex-gateway already forwards arbitrary headers verbatim, so the opt-in works end-to-end without gateway changes. 5. **helexa-acp's `openai_chat` provider** sets `x-include-thinking: true` on every request so its existing `ThinkParser` keeps receiving the marked content stream. `ThinkParser` itself is unchanged — needed for endpoints that aren't reasoning-aware (OpenRouter, OpenAI directly, etc.). ## Acceptance - Zed's commit-message generator (vanilla chat-completions client, no `x-include-thinking`) gets clean commit messages with no `<think>` block. - helexa-acp sessions continue to render thinking in Zed's thought UI via the opt-in path. - Models without reasoning tokens declared in their tokenizer pass through unchanged. - Implementation contains zero references to "qwen3" or any specific model — entirely driven by tokenizer metadata. ## Tests 9 new tests in `wire::event` (token-pair detection across 4 marker conventions, edge cases) and `wire::openai_chat` (default drop, opt-in re-wrap with multi-chunk reasoning, close-marker on Finish, fallback when markers absent, off-switch with markers present). All 213 workspace tests pass; fmt + clippy clean. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-31 17:55:04 +03:00
rob thijssen	957f704efa	feat(neuron): OpenAI Responses API + ci cuda-check runner label ... Step 2 of the Responses rollout: native `/v1/responses` endpoint on neuron that consumes the same InferenceEvent stream as `/v1/chat/completions` but emits it as the Responses API's named SSE event family. No gateway-side translation. ## Surface - `cortex-core::responses` envelope types: `ResponsesRequest`, `ResponsesInput` (text | items), `ResponsesInputItem` (message | function_call | function_call_output | reasoning), `ResponsesContentPart` (input_text | input_image | output_text), `ResponsesResponse`, `ResponsesOutputItem`, `ResponsesUsage`. Plus a `events::*` constant module so the projector and the wire shape stay in sync without string-typos. - `neuron::wire::openai_responses`: - `request_to_chat(req)` flattens Responses input + instructions into a `ChatCompletionRequest` the candle harness already understands. Text-only Parts collapse to a string; mixed text+image Parts go to chat's content-array shape; reasoning items drop; function_call / function_call_output round-trip via tool_calls / tool_call_id metadata so the surface is consistent for the day the harness emits tool calls. - `project_responses_stream(rx, meta)` reads InferenceEvents and emits the eight named events that compose a Responses stream: response.created → output_item.added → content_part.added → output_text.delta×N → output_text.done → content_part.done → output_item.done → response.completed. Synthesises start frames if the producer skips Start (poisoned model, early disconnect) so the stream stays coherent. - `build_response(meta, text, reason, usage)` for the non-streaming path. - `CandleHarness::inference_stream(req)` extracted from `chat_completion_stream`, returning a typed `InferenceStream` (event receiver + id/created/model_id metadata). Both `chat_completion_stream` and the new `responses_stream` are now thin wrappers that pick their wire projection. TP path got the same treatment (`chat_completion_tp_stream` → `inference_tp_stream`). - `POST /v1/responses` route on neuron. Non-streaming returns one buffered `ResponsesResponse`; streaming returns axum SSE with both event names and JSON data per frame (Responses, unlike chat completions, uses named `event:` lines). Reused `inference_error_response` helper hoisted out so the chat and responses handlers share the InferenceError → HTTP mapping. ## CI Also bundles the `cuda-check` runner-label fix from feedback on commit 1859777: `runs-on: rpm` doesn't ship the CUDA toolkit so cudarc's nvcc-version build script blew up. Switched to `runs-on: cuda-13.0` per the existing labels. ## Scope cuts (documented in the modules) - `previous_response_id` rejected at translate time with 400 (`code: chained_conversation_not_supported`) — stateful chained conversations need a persistence layer we haven't built. - Reasoning items dropped (no Qwen3 `<think>` routing yet). - Single output item per response (one `"message"` carrying text); `function_call` items reserved but not synthesised. - Streaming events cover the core set; `response.in_progress` and the web_search / image_generation event families are out-of-scope. 22 new tests: 5 in cortex-core (envelope round-trips), 13 in neuron::wire (request translator + projector + non-streaming builder), 4 in neuron's tests/api.rs (route surface — 503 when no candle, 400 on previous_response_id, 404 on missing model for both stream and non-stream). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-31 11:13:44 +03:00
rob thijssen	6927286cab	fix(neuron): clone id/model_id before TP spawn so wire projector can use them ... The Step 1 refactor moved the InferenceEvent receiver wrap to *after* the orchestration spawn in chat_completion_tp_stream, but the spawn moves both `id` and `model_id` into its async closure (used heavily by acquire_pool_lock, NCCL ops, and tracing). Result: borrowck error E0382 use-of-moved-value on the wire_chat::project_chat_stream call. The non-CUDA build doesn't exercise this branch (it lives behind `#[cfg(feature = "cuda")]`) which is why the workspace clippy/test gate passed locally and on the regular CI workflow. The RPM build workflow, which compiles with --features cuda, caught it (run 244 jobs 2/3/4 against beast / ampere / ada respectively, all the same error). Fix: snapshot `id` and `model_id` into `projector_id` / `projector_model_id` before the spawn, use those at the projector call site. The originals stay free to be moved into the closure. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-31 09:37:10 +03:00
rob thijssen	302ccfb982	refactor(neuron): introduce InferenceEvent + wire projection layer ... Step 1 of the OpenAI Responses API rollout. Pure refactor — no new endpoints, no behaviour change on the wire. Lays the seam for emitting Responses-shaped streaming events from the same harness output as chat completions in Step 2. - New `neuron::wire` module tree: - `wire::event::InferenceEvent` — format-agnostic enum (Start, TextDelta, ReasoningDelta, Finish) the candle harness now emits as its native streaming currency. - `wire::event::FinishReason` — typed reason that maps cleanly onto OpenAI `finish_reason`, OpenAI Responses `status`, and Anthropic `stop_reason` strings. - `wire::openai_chat::project_chat_stream` — async task that consumes an InferenceEvent receiver and produces a ChatCompletionChunk receiver, stamping per-request metadata (id, created, model_id) onto every chunk. Output matches the pre-refactor wire shape bit-for-bit. - candle.rs refactored to emit InferenceEvent on its internal channel through all three streaming paths (CPU run_inference_streaming, CUDA single-GPU stream_inference_via_worker, CUDA TP chat_completion_tp_stream). The streaming functions lost their id/created/model_id parameters since wire-format metadata now lives in the projector. - emit_delta + emit_delta_blocking simplified to single-purpose TextDelta emitters with no wire-format coupling. - chat_completion_stream wraps the InferenceEvent receiver in wire_chat::project_chat_stream before returning so the /v1/chat/completions HTTP handler keeps consuming ChatCompletionChunks unchanged. External signature preserved. Also fixes a pre-existing helexa-acp test race (three modules each declared their own static LOCK for HOME mutation, so cross-module parallelism flaked tests that read HOME at runtime). Consolidated onto a single crate-wide path_util::ENV_LOCK. 122 helexa-acp tests + 44 neuron tests pass (5 new wire projection tests). fmt + clippy --workspace -- -D warnings clean. Ran helexa-acp suite 3x to confirm the env race is closed. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>	2026-05-29 11:30:17 +03:00
rob thijssen	abbedf8d8a	chore(neuron): bump default max_tokens from 512 to 8192 ... 512 is too low for any modern coding model — clients that don't explicitly set max_tokens get clipped responses with no diagnostic. Bump the fallback at all four inference call sites (single-GPU streaming + non-streaming, TP leader + non-leader) to 8192, which fits comfortably within Qwen3-class context windows after a typical agent prompt and lines up with what helexa-acp / a0 / curl clients reasonably expect. Clients that explicitly set max_tokens (now including helexa-acp via HELEXA_ACP_MAX_TOKENS / per-endpoint TOML) override this. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-28 12:38:28 +03:00
rob thijssen	e267f583e1	chore(neuron): rustfmt drift in is_device_fault test ... One assert! call grew past the line limit after the previous commits; cargo fmt --all picked it up. No behavior change. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-28 08:13:55 +03:00
rob thijssen	249b2e5c98	fix(neuron): only poison the model on actual device faults ... Previously every inference Err — shape mismatch, NaN logits, tokenizer error, missing handle — marked the model poisoned and rejected every subsequent request until an operator unload+reloaded. The benjy incident on 2026-05-27 showed how this misfires: a concurrency bug produced a `broadcast_add: shape mismatch` error that had nothing to do with CUDA, but the model was taken down anyway. Add `is_device_fault(err_chain: &str)` — a conservative classifier that returns false only for errors we know are pre-kernel / CPU-side (shape mismatches, NaN logits, tokenize/detokenize, missing handle, DecodeStream, empty prompt). Everything else defaults to true so a genuine driver fault still poisons. Applied at all six poisoning sites: - chat_completion CUDA worker path - chat_completion CPU spawn_blocking path - chat_completion_stream CUDA worker path - chat_completion_stream CPU spawn_blocking path - chat_completion_tp non-streaming wrapper - chat_completion_tp_stream spawned task Each site now logs either "model marked poisoned" (device fault) or "model NOT marked poisoned" (non-device) so the journal makes the classification visible. Tests cover the known non-device patterns and a couple of real CUDA driver messages. Pairs with the inference_lock commit (c59da83): together they eliminate both the cause of the spurious-poisoning we just observed (the shape mismatch) AND the over-reaction to it (the unconditional poison). Each fix is independently useful but the combination is what makes the system actually robust to concurrent agent workloads. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 18:57:48 +03:00
rob thijssen	c59da83636	fix(neuron): serialise single-GPU inference per loaded model ... Two concurrent chat_completion requests against the same single-GPU model could interleave their `clear_kv_cache → forward(chunk0) → forward(chunk1) → ...` sequences. The device-worker channel serialises individual jobs but not the sequence boundary, so the cache could end up holding tokens from one request while another's mask was sized for its own prompt — producing a shape mismatch mid-prefill. Observed on benjy 2026-05-27 18:41:05: agent-zero's `memorize memories` and `memorize solutions` extensions fired 4ms apart against Qwen/Qwen3-8B (a0's utility model). Both prefilled into the same KV cache, and request a08b4a's chunk 0 forward produced scores of shape [1, 32, 512, 1024] against a mask of [1, 1, 512, 512] — broadcast_add failed, both requests bubbled the error up, both flipped the model to poisoned. Add `LoadedModel.inference_lock: tokio::sync::Mutex<()>`, mirroring the TpLoadedModel.pool lock that the TP path already held. Acquire it at the start of `chat_completion` and inside the spawned task of `chat_completion_stream` (so the role chunk goes out immediately and only the inference work queues behind the lock). The CPU branch uses `blocking_lock` from inside spawn_blocking; the CUDA branch uses async `.lock().await` inside tokio::spawn. Throughput impact: zero. The GPU was already serialised at the device-worker channel — multiple requests just produced corrupt KV cache state instead of clean serial throughput. The lock makes the existing serialisation honest. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 18:54:04 +03:00
rob thijssen	f05882369d	fix(neuron): don't poison the model on tokio JoinError panics ... CUDA driver failures propagate as Err through `?` and become `Ok(Err(InferenceError::Other(_)))` from the spawned task — those are real device faults and still poison the model. Tokio JoinError is different: it fires on Rust-level panic (tokenizer bug, sampler bug, serialisation, the UTF-8 slice that landed in commit bd04d7f before the fix) or task cancellation. Those don't touch the device context, so failing the one request without tearing down the model is correct. Two sites changed: - chat_completion's CPU spawn_blocking handler — JoinError no longer sets loaded.poisoned. - chat_completion_tp's tokio::spawn wrapper — JoinError no longer sets tp_for_marker.poisoned. The inner-Err case still does. Each path logs the cause (panicked / was cancelled / ended abnormally) explicitly so the journal makes the new behaviour obvious — search for "model NOT marked poisoned" to find these events. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 18:02:52 +03:00
rob thijssen	bd04d7f580	fix(neuron): stream tokens via DecodeStream to avoid UTF-8 panic ... When BPE byte-fallback splits a multi-byte UTF-8 char (e.g. an emoji) across multiple tokens, the previous "decode the cumulative token list, byte-slice the delta against a stored prefix" pattern would panic with 'start byte index N is not a char boundary; it is inside <emoji>'. The race: at step N the tokenizer renders the partial bytes as U+FFFD (3 bytes); at step N+1 it can decode the complete codepoint (e.g. 4 bytes for 🌫). `decoded_prefix.len()` from step N then lands inside the codepoint in step N+1's `full` string, and `&str[start..]` panics. Replace with tokenizers' `DecodeStream::step(id)` which maintains an internal byte buffer across token boundaries and only emits when a clean codepoint completes. Applied at all three SSE emission sites: - stream_inference_via_worker (single-GPU CUDA stream) - chat_completion_tp_stream's spawned task (TP stream) - run_inference_streaming (CPU stream) The shared emit helper splits into emit_delta (async, mpsc::send) and emit_delta_blocking (sync, mpsc::blocking_send) so each path keeps its existing send semantics. The old emit_chunk helper that did the unsafe full-decode-and-slice is removed entirely. Observed on beast 2026-05-27 17:49:55 — model emitted 🌫 in a tool-call response after a long agent-zero session; the spawned TP stream task panicked at candle.rs:2648. The model itself stayed healthy (no CUDA fault), only the one streaming request died. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 18:01:24 +03:00
rob thijssen	1e13889392	feat(neuron): chunked prefill + VRAM/prompt-length pre-flight checks ... Prevents the OOM-during-prefill → poisoned-context → 5-minute-reload cycle observed on beast under agent-zero workloads. Three changes, all keyed off env-driven knobs so an operator can tune without a rebuild: 1. Chunked prefill (NEURON_PREFILL_CHUNK_TOKENS, default 512). The initial forward is split into N-token windows, each with a monotonically growing offset. KV cache accumulates across chunks exactly as it would under one big prefill; only the final chunk's logits are kept for sampling. Activation memory now scales with chunk size instead of prompt length, so a 13 k-token prompt stops holding tens of GB of intermediate activations live at once. Wired into all six prefill call sites: - run_inference / run_inference_streaming (CPU path) - run_inference_via_worker / stream_inference_via_worker (CUDA single-GPU through device worker) - chat_completion_tp_inner / chat_completion_tp_stream (TP via WorkerPool) Three helpers — chunked_prefill_local, chunked_prefill_via_worker, chunked_prefill_tp — own the loop shape so the chunking semantics stay identical across paths. Per-chunk debug log shows progress. 2. Max prompt length (NEURON_MAX_PROMPT_TOKENS, default 16384). Requests above the cap return a structured 400 with `code: prompt_too_long` rather than going through the prefill and discovering the limit by OOMing partway through. New InferenceError::PromptTooLong variant. 3. Minimum free VRAM gate (NEURON_MIN_FREE_VRAM_MB, default 1500). If `vram_free_mb` is below the threshold at request start (e.g. another concurrent request is mid-prefill), reject with a clean 503 + `code: insufficient_vram` rather than starting work that will OOM. New InferenceError::InsufficientVram variant. CPU loads (vram=0 sentinel) skip this check. All three gates fire BEFORE any device work, so a rejected request costs ~one tokenisation pass and never touches the worker thread — poison cascades from rejected work are now impossible. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 13:46:54 +03:00
rob thijssen	35876954cd	chore(neuron): default tracing filter to info (was info,neuron=debug) ... Production deployments that want neuron-internal debug detail (e.g. trim_device_pool's per-clear-kv line, slab inserts/drops) override RUST_LOG explicitly via systemd. Defaulting to debug for the whole neuron target produced a lot of journal volume that wasn't useful in the common case. beast already sets RUST_LOG=debug in /etc/systemd/system/neuron.service.d/local.conf, so beast's verbosity is unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 12:47:30 +03:00
rob thijssen	cdf0f4e66d	fix(neuron): trim cudarc mempool after clear_kv_cache to release VRAM ... cudarc's stream-ordered memory pool retains freed blocks (cuMemFreeAsync returns memory to the device's default mempool, not to the OS), so mem_get_info under-reports free VRAM between requests. With Qwen/Qwen3.6-27B TP=2, the second consecutive chat completion saw ~4.5 GB of "missing" free VRAM and either OOMed or tripped cuBLAS into CUBLAS_STATUS_INTERNAL_ERROR depending on quant. Add a cuda-gated trim_device_pool helper that, after each successful clear_kv_cache, synchronizes the context and calls cuMemPoolTrimTo(pool, 0) against the device's default mempool. Failures (no async-alloc support, transient driver errors) are non-fatal and log at debug. The before/after free-VRAM delta is logged so an operator can correlate the trim with the next request's prefill VRAM. ConcatKvCache::reset() in candle-nn 0.10.2 already drops its tensors correctly; the leak was strictly at the cudarc pool layer. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 12:36:13 +03:00
rob thijssen	b4f3576d82	refactor(neuron): phase 4 — model loads move onto the device worker ... Final structural slice of the per-device CUDA context-ownership refactor. The four remaining spawn_blocking sites that did CUDA work on the leader are gone: - Single-GPU GGUF load (`load_arch_gguf` spawn_blocking) → `Job::LoadGguf` dispatched on the worker. - Single-GPU dense load (`load_arch_dense` spawn_blocking) → `Job::LoadDense` on the worker. - TP shard load (`WorkerPool::load_dense_shard` spawn_blocking) → `Job::TpLoadShard`. The dispatch handler reads `state.nccl.comm()` directly — no cross-thread `Arc<Comm>` transfer, no `SendComm` wrapper for this path. The Phase 2 / Phase 3 bridges that moved freshly-built models across the channel boundary (`Job::TransferIn`, `Job::TransferInTp`, `Job::CloneLeaderComm`) are removed. Models are now constructed on the worker thread directly; the slab gets populated by `insert_arch` / the inline `tp_models.insert` in dispatch handlers. What this phase preserves: - CPU loads still use `tokio::task::spawn_blocking` against `Arc<Mutex<ModelArch>>`. There's no CUDA context to own on CPU and channel overhead would only add latency. Four `spawn_blocking` references remain in `candle.rs` (load_arch_gguf, load_arch_dense, chat_completion, chat_completion_stream) and all are deliberate CPU-only fallback. - Public API unchanged. `Harness::load_model`, `chat_completion`, HTTP routes all keep identical signatures. What this phase removes: - `SendComm` wrapper is no longer used in the load path (the Phase 3 bridge that justified it). It remains in `nccl_state.rs` for the Phase 1–3 era and any future cross-thread Comm move; consider deleting in a follow-up. - `Job::TransferIn`, `Job::TransferInTp`, `Job::CloneLeaderComm` and their handle convenience methods deleted. - The leader_device parameter on `load_dense_shard` is now `_` — unused since the worker has its own bound device. Removing the arg outright is a public-API change; keeping the underscore prefix preserves the signature and signals deadness without churn. Helper relocation: - `LlamaDense::from_parts` is a new pub(crate) constructor so the worker-thread loader can build a `LlamaDense` without going through the original `load_arch_dense` async function. - `check_dense_config_supported` is bumped to `pub(crate)` for the same reason. Sweep verified: `grep -rn spawn_blocking crates/neuron/src/harness/` returns only CPU-fallback hits in `candle.rs` + doc-comment references to the old design. All four leader-side CUDA `spawn_blocking` sites are gone. fmt + clippy clean; 37 lib tests + all integration tests pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 10:24:38 +03:00
rob thijssen	76ab24d98c	refactor(neuron): phase 3 — TP forward + NCCL state move onto device worker ... Third slice of the per-device CUDA context-ownership refactor planned at ~/.claude/plans/plan-the-per-device-worker-abstract-micali.md. The leader's `NcclState`, every `Comm::all_reduce` issued by the TP layers, the leader-side KV cache reset, and the TP forward step itself now all run on the per-device worker thread — the same OS thread that bound the leader's `CudaContext` at startup. What this phase changes: - `Job` gains `NcclInit`, `NcclSanity`, `CloneLeaderComm` (Phase 3 bridge — Phase 4 removes), `TransferInTp`, `DropTp`, `TpClearKv`, `TpForwardLogits`. Plus a new `TpHandle(u64)` opaque key. - `DeviceWorkerState` gains `nccl: NcclState` and `tp_models: HashMap<TpHandle, Box<TpLeaderModel>>` (+ counter). - `WorkerPool` loses its `leader_nccl` field; gains a `leader_worker: Arc<DeviceWorkerHandle>` passed at construction. `init_nccl`, `nccl_sanity_check`, `load_dense_shard`, `generate_step`, `clear_kv_cache` all route their leader-side ops through `Job::Nccl*` / `Job::Tp*` instead of spawn_blocking against a Mutex-wrapped state. `generate_step` returns `Vec<f32>` instead of a device-resident `Tensor` — the worker copies logits to CPU before reply so the async caller can sample on a CPU candle tensor with zero device-context touch. - `TpLoadedModel.leader_model: Arc<Mutex<TpLeaderModel>>` → opaque `leader_handle: TpHandle`. The boxed `TpLeaderModel` lives in the worker thread's slab; both the model's CUDA tensors and the embedded `Arc<Comm>` clones release on the same thread that allocated them (the Drop semantics constraint cudarc forces). - `Job::CloneLeaderComm` is a Phase 3 bridge: the TP shard load still runs in spawn_blocking and needs the leader's `Arc<Comm>` to build the row-parallel layers' AllReduce ops. The Job clones the Comm out of the worker's NcclState and ships it back as `SendComm`. Phase 4 deletes this bridge when the load itself moves onto the worker. - `Job::NcclInit` and `Job::NcclSanity` are ungated by `cuda` so the no-cuda `NcclState` stubs (which reply with `cuda_feature_not_enabled`) still flow through the same channel uniformly; the cuda-only TP variants (CloneLeaderComm, Transfer/Drop/Clear/Forward Tp) remain gated. What this phase doesn't touch (yet): - TP shard load itself — still spawn_blocking, bridged via `CloneLeaderComm`. Phase 4 moves it to `Job::TpLoadShard` and reads `state.nccl.comm()` directly inside the worker. - Single-GPU model loads — still spawn_blocking, transferred via `Job::TransferIn`. Phase 4 moves them. - `device_vram_mb` / `cuda_mem_mb` / `log_construction_complete` helpers — still present, used inside spawn_blocking load closures. Phase 4 cleanup folds them into `dispatch.rs`. `tp/mod.rs::WorkerPool::spawn` gained a required `leader_worker: Arc<DeviceWorkerHandle>` argument. Three external callers were updated: `CandleHarness::load_tp` (passes the cached device worker), `main.rs::tp_smoke` (spawns a fresh worker), and the two `tp_worker_lifecycle*.rs` integration tests. Public API unchanged. fmt + clippy clean; 37 lib tests + all integration tests pass. CUDA-only TP integration smoke deferred to the next deploy on beast. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 10:16:02 +03:00
rob thijssen	b179204fd3	refactor(neuron): phase 2 — single-GPU forward + clear_kv route through device worker ... Second slice of the per-device CUDA context-ownership refactor planned at ~/.claude/plans/plan-the-per-device-worker-abstract-micali.md. The two spawn_blocking sites in `chat_completion` and `chat_completion_stream` now route through the device worker thread on CUDA loads. CPU loads keep the existing spawn_blocking + `Arc<Mutex<ModelArch>>` path; there's no context to own and the channel hop would only add latency. What this phase changes: - `Job` gains `TransferIn`, `DropArch`, `ClearKv`, `ForwardLogits`. The worker's dispatch state grows a `HashMap<ArchHandle, Box<ModelArch>>` slab and a `next_handle` counter for minting opaque handles. - `LoadedModel.arch: Arc<Mutex<ModelArch>>` → `Option<Arc<Mutex<>>>`, plus a new `arch_handle: Option<ArchHandle>` field. The two are mutually exclusive: CUDA loads set `arch_handle = Some(_)` after transferring the boxed arch into the worker's slab; CPU loads keep `arch = Some(_)` for the legacy spawn_blocking path. - New `run_inference_via_worker` and `stream_inference_via_worker` drive the prefill + decode loop by sending `Job::ForwardLogits` per step; the worker copies the resulting `[vocab]` logits to a CPU-side `Vec<f32>` before reply, so the async caller never holds a device-resident tensor. `apply_repeat_penalty` and `LogitsProcessor::sample` run on a CPU candle tensor; no context binding side-effects on tokio worker threads. - `logits_health_slice(&[f32])` complements the existing `logits_health(&Tensor)` so the new worker paths can compute health stats directly from the CPU vec. - `unload_model` for the single-GPU CUDA path now sends `Job::DropArch { handle }` to the worker so the `Box<ModelArch>` drops on the thread that allocated its CUDA tensors. The `Drop` runs with the bound context, freeing memory on the right context. What this phase doesn't touch (yet): - TP forward, TP load, NCCL bring-up — still on spawn_blocking. Phase 3. - Single-GPU model load — still spawn_blocking, followed by a `Job::TransferIn` to move the freshly-built `ModelArch` into the worker slab. Phase 4 moves the load itself onto the worker thread and eliminates the bootstrap TransferIn. - The `device_vram_mb` / `cuda_mem_mb` helpers — still present and used by the construction-time logs running inside spawn_blocking loads. Phase 4 cleanup folds them into `dispatch.rs`. Public API unchanged. fmt + clippy clean; 37 lib tests + all integration tests pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 09:55:08 +03:00
rob thijssen	081b532387	refactor(neuron): phase 1 — per-device worker thread, VRAM queries route through it ... First slice of the per-device CUDA context-ownership refactor planned at ~/.claude/plans/plan-the-per-device-worker-abstract-micali.md. Adds the infrastructure for a dedicated OS thread per CUDA device that owns the device's `CudaContext` for the daemon's lifetime, and routes the 8 async-context `device_vram_mb()` call sites in candle.rs through it. What this phase changes: - New module `harness/device_worker/` (mod.rs, jobs.rs, dispatch.rs). `DeviceWorkerHandle::spawn(idx)` creates a named OS thread (`cuda-dev-N`), binds `CudaContext::new(idx)` once at startup, and enters a dispatch loop reading `Job`s off a `std::sync::mpsc` channel. Replies cross back via `tokio::sync::oneshot::Sender` so async callers await without parking a tokio worker. - Two Job variants: `QueryVram` and `Shutdown`. Phases 2–4 add Forward, ClearKv, NCCL init/sanity, and load variants. - `LoadedModel` and `TpLoadedModel` gain a `worker` field populated at load time by a new `CandleHarness::ensure_device_worker(idx)` method that lazily spawns + caches one worker per device index. - Per-model `query_vram()` convenience method on both struct types so the 8 call sites in chat_completion / chat_completion_stream / chat_completion_tp_inner / chat_completion_tp_stream become `loaded.query_vram().await` (or `tp.query_vram().await`) — same field values logged, just sourced from the owner thread instead of the caller thread. What this phase doesn't touch (yet): - Forward, kv-cache clear, model load, NCCL — still on `spawn_blocking`. Phase 2 moves the single-GPU forward + clear; Phase 3 moves the TP forward + NCCL bring-up; Phase 4 moves the loads and deletes the now- unused `device_vram_mb` / `cuda_mem_mb` helpers. - Public API — unchanged. `Harness::load_model`, `chat_completion`, HTTP routes all keep identical shapes. Tests: - 5 new unit tests in `device_worker/mod.rs::tests` cover spawn → query → shutdown round-trip, thread naming, post-shutdown submit returns `Gone`, poisoned flag fast-rejects, and concurrent jobs drain across a Shutdown. CPU build (the only one CI runs) is enough to exercise channel mechanics. - All 37 lib tests + all integration tests pass; fmt + clippy clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 09:40:34 +03:00
rob thijssen	7c19da9361	feat(neuron): construction-complete vram/config dump + logits health + per-step vram ... Three additive diagnostics that turn the 2026-05-27 q5k Qwen3.6-27B incident from "guess at KV cache / quant sizes" into "read the journal": 1. Construction-complete summary in TpQwen3_5ForCausalLM::load and TpQwen3ForCausalLM::load. After the last "after layer N" log fires, each rank emits a single info line with: free_mb/total_mb (the number that drops by ~9 GB between per-layer and first-request on beast, with no inference traffic), every resolved config knob (vocab_size, hidden_size, num_layers, head_dim, num_kv_heads, max_position_embeddings), and a per-token KV-cache byte estimate. For Qwen3-Next also includes the linear/full-attention layer split so the hybrid architecture's cache cost is unambiguous. 2. Logits health snapshot on sample failure. Today the failure logs "A weight is negative, too large or not a valid number" with no context — was it a NaN cascade, an Inf, a negative weight? `logits_health(&logits)` computes nan/pos_inf/neg_inf/neg counts plus finite_min/max/mean on the failure path (zero cost on the success path) and emits a warn line just before the wrapper's terminal "failed, model marked poisoned" log. Wired into both the prefill and decode sample sites of the non-streaming AND streaming TP chat paths. 3. VRAM snapshot at prefill complete + every decode step. The "prefill complete" info line now carries vram_free_mb so the activations + KV growth from the prefill itself is visible. The per-step trace line gets vram_free_mb too, so an operator running with RUST_LOG=trace can watch headroom shrink token by token. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-27 09:04:55 +03:00
rob thijssen	800498f530	feat(neuron): bind listener before pre-warm, surface activation in /health ... Two coupled changes addressing the 2026-05-26 validate-neuron failure where a fresh deploy of beast had /health unreachable for ~5 minutes while Qwen3.6-27B q5k materialised, even though systemd reported the unit as active. 1. main.rs no longer awaits load_default_models before binding axum. The listener binds first; pre-warm runs in a spawned background task that holds a read lock on the harness registry for the duration of its sequential load loop. Concurrent on-demand /models/load and /v1/chat/completions traffic still flow. 2. /health gains an `activation` field carrying: state pre_warming | ready pending model ids queued but not started in_progress model id currently loading (Option) completed model ids loaded successfully this activation failed [{model_id, error}] for failed entries The field is `#[serde(default)]` so a pre-change cortex polling a new neuron — or vice versa — keeps working. `ActivationTracker` (new module `neuron::activation`) owns the RwLock-wrapped state; load_default_models takes a tracker reference and updates it per-model. NeuronState holds an Arc clone for the /health handler. Tests updated to construct trackers and assert state transitions (empty noop, two failures → ready with both in `failed`). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-26 15:18:04 +03:00
rob thijssen	2740e61a23	fix(neuron,candle): name lifetime on acquire_pool_lock ... Lifetime elision fails when a function has two reference parameters and returns a borrow: rustc can't infer whether the MutexGuard's lifetime ties to `pool` or `model_id`. The non-CUDA build skipped this code path (cfg-gated), so the error only surfaced on the GPU build at https://git.lair.cafe/helexa/cortex/actions/runs/162. The guard borrows the pool, so name the lifetime on `pool` and the return type. `model_id` keeps its independent (elided) lifetime. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-26 12:37:32 +03:00
rob thijssen	67f79c868f	fix(neuron,shutdown): time-bound unloads, fast-exit past tokio drain ... Two failure modes from the 2026-05-26 beast incident: 1. `unload_all_models` looped through models calling `unload_model`, logging individual failures at warn. The cumulative effect was a single warn line for the failed unload then "shutdown complete" — no signal that the model was actually still loaded. Now each unload is bounded by a 20s timeout, failures escalate to error, and a summary "leaving N model(s) loaded" line fires when anything is stuck so the operator knows the OS will reclaim VRAM after exit. 2. Returning `Ok(())` from `main` after the unload sweep dropped the tokio runtime, which then waited indefinitely on a CUDA-stuck spawn_blocking thread (the journal's "Stack trace of thread 2951308" — spinning on `cuCtxGetCurrent`). systemd's TimeoutStopSec fired 2 minutes later, SIGABRT, core dump. Replacing the return with `std::process::exit(0)` skips the runtime drain and hands the OS a clean exit code; stuck threads get reaped with the process. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-26 12:30:06 +03:00
rob thijssen	fc6ef0ee0f	feat(neuron,candle): detect CUDA context poisoning and refuse follow-ups ... Once a CUDA driver error has hit a forward or kv-cache call, the device's context is unrecoverable in-process — subsequent kernels can hang (the failure mode seen on beast on 2026-05-26), return garbage, or trip another illegal-address. The harness now marks the model poisoned on any forward / spawn_blocking / TP-task failure, refuses further inference against it with a clear "unload and reload" error, and surfaces `status: "poisoned"` on `/models` so an operator running `curl beast:13131/models` (or cortex polling) can see the bad state. Without this, a single OOM on a too-large prefill quietly turned every subsequent request into a stuck wait on the pool lock; with it, the first request fails fast with the driver error in the journal and the client gets a usable 5xx instead of a hung connection. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-26 12:28:42 +03:00
rob thijssen	1385979e3d	feat(neuron,candle): log per-device VRAM at chat_completion start ... Every "starting" log line now carries vram_free_mb / vram_total_mb for the request's serving device (the leader device on TP). On the 2026-05-26 incident this would have made the 14k-token prefill OOM diagnosable from the first log line: with ~412 MB free, that prompt was never going to fit, and the operator could have caught the imbalance before the CUDA context got poisoned. `device_vram_mb` mirrors the existing helper in tp_qwen3_5.rs and is kept separate to avoid coupling the inference path to the TP module. TpLoadedModel gains a `leader_device: Device` clone so the request path reads the device without locking the leader model (which would contend with an in-flight forward). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-26 12:26:23 +03:00
rob thijssen	0a1cfcd4d0	feat(neuron,candle): req_id spans, terminal failure logs, pool-lock warnings ... Every chat completion path (single-GPU + TP, streaming + non-streaming) now opens an `info_span!("chat", req_id=…, model=…)`. The fmt subscriber prefixes every event with that span so `grep req_id=…` over journalctl reconstructs one request even when dozens overlap. Every path also emits a terminal log line on both success ("done", with prompt_tokens/completion_tokens/finish_reason/total_ms) and failure ("failed", with full anyhow chain + total_ms). Failures used to vanish silently — a request that hit a CUDA OOM left "starting" in the journal and no further trace. New `acquire_pool_lock` helper replaces the bare `tp.pool.lock().await` in both TP paths. It warns at 2s ("still waiting on pool lock") and re-warns every 2s thereafter, so queued requests stuck behind a deadlocked holder are visible immediately instead of looking like idle silence. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-26 12:25:11 +03:00
rob thijssen	ea0e0f7911	fix(neuron,tp): log leader forward errors with full context ... Worker rank failures were already surfaced at WARN, but the leader's own forward Result::Err was silently coerced to a `leader_ok=false` bool. When the leader and a worker both fail together — the typical shape of a CUDA OOM cascading into an illegal-address — the journal showed only the worker side and an operator had to guess what hit rank 0. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-26 12:22:30 +03:00
rob thijssen	e71181499e	feat(stage-8e-3): quantize lm_head in TP Qwen3-Next ... TpQwen3_5ForCausalLM::lm_head is now a MaybeQuantLinear. When the load spec has quant set and tie_word_embeddings is false, lm_head's (vocab_size, hidden_size) weight is quantized in-situ at load time along with all the per-layer linears. The non-tied case on Qwen3.6-27B saves ~1.7 GB per rank vs bf16 (248320 x 5120 x 2 bytes = 2.42 GB -> ~700 MB at Q5K) and shaves a small amount of decode latency from the per-token logits matmul. Tied case (tie_word_embeddings=true) keeps the lm_head plain even when quant is set — quantizing the shared tensor would corrupt the embedding lookup, and the tied case already gets the memory win from only holding one copy. This is the last MaybeQuantLinear hookup in the Qwen3-Next TP path. The dense Qwen3 path (tp_qwen3.rs) is unchanged — defer until it's the bottleneck for a model that actually needs TP at consumer scale. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-21 21:53:14 +03:00
rob thijssen	ee663e5e99	fix(stage-8e-2e): bump quant prefill threshold to M > 64 ... The M > 8 threshold from 8e-2d activated forward_via_f16 on the test case (M=30) and slightly regressed prefill (143 -> 133 T/s). The dequant cost (~30 MB f16 per linear * ~480 calls per prefill = ~200 ms) eats the cuBLAS GEMM speedup at small M. Move the crossover to M > 64 so short prefills (typical for the validate probe) stay on the GGUF GEMV kernel where per-call cost is comparable but the dequant tax is zero. Long prefills still get the dequant-then-cuBLAS-GEMM path where the GEMM scaling amortises the fixed dequant cost. Doesn't close the gap to mistralrs's 423 T/s on Q5K prefill — that needs either a dequant cache (gives back the ISQ memory win) or a fused dequant+gemm kernel. Both larger projects. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-21 21:50:45 +03:00
rob thijssen	34f9b77d9d	feat(stage-8e-2d): route quantized matmul by M (prefill vs decode) ... MaybeQuantLinear::forward picks between two QMatMul paths: - M > 8 (prefill): QMatMul::forward_via_f16 dequantises the weight once into f16 and runs a real cuBLAS-backed GEMM. The dequant cost is fixed per call, so it's amortised across the M tokens. - M <= 8 (decode): QMatMul::forward uses candle's GGUF GEMV kernel on the quantized blocks directly. Requires f32 inputs so we still cast in/out at the boundary in that arm. Earlier 8e-2c sent everything through the GGUF GEMV kernel, which is excellent at GEMV (decode) but doesn't have a real batched GEMM path — prefill regressed ~4x. This restores prefill to roughly the bf16 cuBLAS GEMM throughput while keeping the decode gain. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-21 21:15:32 +03:00
rob thijssen	f084aaab8e	fix(stage-8e-2c): cast bf16/f16 activations to f32 around QMatMul ... candle's QTensor::cuda_fwd requires f32 inputs — its on-the-fly GGUF dequantize accumulates in f32. The model dtype flowing into MaybeQuantLinear::forward is bf16, so QMatMul::forward errored with "unexpected dtype, expected: F32, got: BF16". Wrap the Quant arm to cast the activation to f32 before the matmul and cast the result back to the input dtype. The cast is a single launch on the activation tensor (small relative to weight traffic); it's the price of in-situ GGUF-style quantization, and what mistralrs does inside its own Linear wrapper. The Plain arm is unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-21 20:05:19 +03:00
rob thijssen	68a606a79c	fix(stage-8e-2b): allow quant on the TP load path ... The pre-existing guard in candle.rs rejected any spec.quant on the TP path with "GGUF quantized models are not supported in the TP path" — written when quant only ever meant GGUF. With 8e-1/8e-2 in, quant != None on the TP path triggers in-situ quantization of the loaded safetensors shards. resolve_dense_files only looks for safetensors so a GGUF-source-file model with TP still errors out cleanly downstream. validate-neuron.sh: rebuild the load payload incrementally so tp_size > 1 + non-empty quant produces both fields. Same script now covers all four combos (single/TP × dense/ISQ). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-21 19:17:14 +03:00
rob thijssen	4aa71902d0	feat(stage-8e-2): plumb quant config from ModelSpec to TP load path ... - LoadDenseShard RPC gains an optional `quant` string field. - WorkerPool::load_dense_shard takes a `quant: Option<String>`, passes it via the RPC to workers and via parse_quant_string to the leader's local load. - The Qwen3-Next TP load chain (ForCausalLM → Model → DecoderLayer → Attention / GatedDeltaNet / MLP) takes `quant: Option<GgmlDType>` end-to-end, calling Column/RowParallelLinear::load_with_quant. - The fused in_proj_qkv inside TpQwen3_5GatedDeltaNet is now a MaybeQuantLinear so it also picks up quantization. - parse_quant_string accepts q4_0/q4_1/q5_0/q5_1/q8_0/q8_1, q2k..q8k (with or without underscore), and f16/bf16/f32. Empty / None means no quantization. Callers from candle.rs forward spec.quant through pool.load_dense_shard. This means a `quant = "q5k"` in models.toml now flows end-to-end to a QTensor-backed QMatMul for every per-rank linear in the Qwen3-Next TP path. Leaves lm_head and the small replicated bias/log tensors in their loaded dtype (Stage 8e-3). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-21 18:03:36 +03:00
rob thijssen	bef159b21c	feat(stage-8e-1): MaybeQuantLinear primitive + parallel-linear quant variants ... Introduces MaybeQuantLinear, which wraps either a plain candle Linear or a candle QMatMul backed by a freshly-quantized QTensor. Forward dispatches identically through the Module trait so downstream code doesn't care which arm is active. ColumnParallelLinear and RowParallelLinear gain `load_with_quant` methods. The existing `load` methods stay as backward-compatible no-quantization wrappers — no churn at the 27 existing call sites. This is the foundation for in-situ quantization at load time. Wiring the user-facing quant config and switching call sites to load_with_quant follow in stages 8e-2 / 8e-3. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-21 17:55:26 +03:00