- 29 Jun, 2026 8 commits
-
-
Stefy Lanza (nextime / spora ) authoredA whole-match regen (scope=full) now exposes two checkboxes on the match detail page that drive the finalization tail of the pipeline: - "Finalize & package" (default on): after assembly, run the 2x AI upscale + 2x frame interpolation pass, generate arbitrage-safe odds, and pack the renamed upload ZIP (each of the nine slots picked from its highest-quality variant via _best_variant). - "Upload after" (default = upload_after_render): also push the prepared match to the configured Township endpoint. Implies package; refuses cleanly (no server call) when the endpoint isn't configured. The full job is now seven phases (was four): prompts, keyframes, render, assemble (always) + enhance, odds/ZIP, upload (gated). The enhance pass was previously always-on; it now sits under the package checkbox (default checked, so existing behaviour is preserved) — unchecking gives a fast base-res iteration with no upscale/odds/zip. Wiring: added package/upload to the /matches/render params allowlist; reMatch() reads the checkboxes for scope=full, folds them into the confirm dialog, and posts them; odds/ZIP via prepare_match_odds_zip and upload via upload_prepared_match (progress piped into the job bar). Co-Authored-By:Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RdMufYvtTbtGDWsiZVoXce
-
Stefy Lanza (nextime / spora ) authored
The video pipeline is reused across clips (model CPU offload keeps weights in host RAM), so the per-request teardown that trims the heap — _free_pipeline_vram -> _trim_cpu_ram — never runs between clips. Each generation allocs ~1 GB of decode/latent/offload buffers; Python frees them but glibc keeps the pages in its arena, so RSS drifts up clip-by-clip (seen ~34 -> 47 GB) and peaks over the max_ram_gb cap. The cap can't reclaim it: the only resident model is the protected live one, so there's nothing idle to evict (and ram_leak_watch is off). Fix: after each successful generation, drop the decode buffers (frames/frame_np), gc, and malloc_trim the freed heap back to the OS. Logs RSS before/after so a RESIDUAL reference leak (RSS not returning to baseline after trim) is visible to chase separately. Failure paths already trim via _free_pipeline_vram. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
manager: universal footprint rule — already-quantized (cached) models measured as-is, no quant factor Per directive: the VRAM estimate must always equal the model's ACTUAL full pipeline footprint, by a universal rule rather than per-model tuning. The quant factor exists to convert an UNQUANTIZED source's weight size down to its quantized runtime size — so it must NOT be applied to weights that are ALREADY quantized on disk. Add _quantized_cache_gb(): find this model's quantized pipeline cache by MODEL NAME (glob '<safe_name>__*', ignoring in-progress '.building' dirs; signature suffix drifts with trivial config edits, so name-match is the robust signal) and return its real on-disk size. _get_model_used_vram_gb now, right after the forced-measurement check, returns that cache size AS-IS (+ reserve) when present — no quant_mult, no precision factor. The cache holds already-quantized weights, so its size IS the footprint, and it overrides a stale used_vram_gb (e.g. Wan2.2-I2V's 151 GB fp32 disk size). Models quantized AT LOAD (no cache yet) still fall through to the factor-based estimate. Verified: VACE-Fun now estimates ~24 GB (real 4-bit footprint) instead of 24.5 x 0.283 = 6.9 GB, so eviction frees the idle image model before the forward. Reverted the per-model measured_vram_gb pin in models.json — the universal rule covers it. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Root cause of the persistent ~11 GB VRAM "base" and the big-clip OOMs: the video model's persisted measured_vram_gb was 0.296 GB — the tiny GPU-resident slice of an OFFLOADED load. With force_vram_update it overrode used_vram_gb=24.5, so ensure_vram_for(video) thought the model needed ~0 GB and NEVER evicted the idle Z-Image image model (~15 GB). That left ~11 GB resident; the video forward then needed ~10 GB on top → 21 GB → OOM. (The whole GPU is coderai's — that base was our own un-evicted model.) Fix (per directive): measured_vram_gb now represents the FULL model footprint — GPU-resident weights + the portion offloaded to host RAM + runtime reserve — i.e. placement-independent total memory the model needs. record_vram_delta computes it from the GPU delta PLUS the host-RAM delta (video path now passes ram_before), and force_vram_update persists that full number every run. So eviction/strategy provision for the real need and free room (evict the idle image model) for the forward. Also cleared the stale 0.296 from models.json so it falls back to used_vram_gb=24.5 until re-measured. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
From the latest render log: the per-component cache + same-model reuse worked (clip 02 ran on clip 01's resident pipe, LoRA-synced, no reload), but BIG clips (46-50 frames) OOM'd in the dual-expert forward under 'model' CPU offload, and the single retry then crashed two ways: - NameError: loading_task was imported only in the load-only block, so on a REUSED pipe the retry's `with loading_task(...)` was unbound. - enable_sequential_cpu_offload on injected bnb-4bit components hits "Params4bit.__new__() got an unexpected keyword argument '_is_hf_initialized'". So every big clip returned 500 instead of recovering. Fixes: 1. Generation fallback LADDER. On a recoverable failure (OOM / device-mismatch) we no longer return 5xx — we free the pipe (outside the except, so the failed forward's traceback can't pin its VRAM) and reload with the next strategy, then retry generation. Return on the FIRST success, or 500 only after EVERY rung fails. Rungs: [already-loaded] -> balanced -> disk, all PLAIN device_map builds (incremental=False) so device_map places every component itself — sidestepping the injected-component failure modes ('model' OOM, 'sequential' Params4bit, device-map-can't-place-injected). loading_task imported up-front. 2. Monolithic cache finalize. After the per-component heavy build, complete the cache DIR (save light components + model_index.json + marker) so later loads use the monolithic HIT path: from_pretrained(dir, device_map=balanced), NO injection — big-clip generation fits AND loads pre-quantized from cache (fast, no re-quantize). Conservative: only marked complete when every component is present. Net: first clip builds+finalizes the cache; thereafter loads are balanced-from-cache (fits any clip size, no injection bugs), same-model clips reuse, and a transient OOM walks the ladder instead of failing the clip. Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
Belt-and-suspenders for the "untracked in-use VRAM" failure class (a teardown that drops a model's registry entry but leaves GPU tensors resident — e.g. activations pinned by an exception traceback): - manager.sweep_orphan_vram(): gc.collect() + empty_cache() reclaims the UNREFERENCED kind (a just-dropped-but-not-collected pipeline), and logs a diagnostic when the card holds materially more VRAM than tracked models account for (the referenced kind, which can't be freed here — must be fixed at source). - manager._estimate_tracked_vram_gb(): sums tracked models' footprints for that check (offloaded models over-report, so the check is conservative, never a false positive on a legitimately offloaded model). - video: call it at the start of every load — right after request_model's eviction — so a model swap is verified to have actually freed the outgoing model's VRAM, and any stray orphan is swept before the new model loads. Same-model consecutive requests already reuse the resident (offloaded) pipe via request_model, so this only runs on a genuine (re)load. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Follow-up to the generation-OOM leak fix: forcing 'sequential' was treating the symptom. The real cause of "model offload OOMs generation" was that the card already had ~11 GB LEAKED from the prior clip's failed forward (its exception traceback pinned the activations). model offload itself adds ~0 to the GPU (weights stay on CPU); subtract the leak and the forward fits in 24 GB. With the leak fixed (free + retry now happen outside the except, so the pinned activations are collectable), loads start from a clean card, so force the faster 'model' CPU offload for injected components instead of sequential. The leak-free retry still degrades to sequential only if a genuinely clean-card forward OOMs. Why load-time cleanup couldn't fix it on its own: the leaked VRAM was neither a tracked model (already popped from the registry) nor free-able (still referenced by the traceback), so manager eviction had nothing to evict and empty_cache() — which only reclaims unreferenced memory — was a no-op. The only fix is dropping the reference at its source, which the traceback fix does. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
video: fix generation-OOM VRAM leak (traceback pin) + use sequential offload for injected components Two issues from the latest debug.log on the cached VACE model: 1. Generation-OOM VRAM leak (the cascade driver). The OOM/retry handler reloaded INSIDE `except Exception as e:`, so sys.exc_info() kept the failed forward pass's traceback alive — pinning its GPU activations (and the resident expert). _free_pipeline_vram() then reclaimed nothing, the retry reload saw the card still ~21 GB full, and every fallback (sequential/disk) OOM'd in turn -> 500. Same traceback-pinning bug fixed earlier for the load ladder. Restructured: capture the error, drop e.__traceback__, and do the free + gc + empty_cache + retry OUTSIDE the except block, where the activations are finally collectable. 2. model CPU offload OOMs this dual-expert forward. With injected cached components we forced 'model' offload, which keeps a whole ~7 GB expert resident; the A14B forward then spiked past 24 GB (loaded at 11 GB, OOM mid-generate). Force 'sequential' instead — minimal footprint that reliably fits; the cache still makes the LOAD fast. (offload_strategy=group remains available for more speed.) Retry policy unified: OOM -> sequential; device-mismatch -> incremental cache off (plain device_map build hooks everything itself). Both retry once, then 500. Co-Authored-By:
-
- 28 Jun, 2026 4 commits
-
-
Stefy Lanza (nextime / spora ) authored
The incremental per-component cache built and cached all components correctly, but generation then failed: RuntimeError: Expected all tensors to be on the same device, but got index is on cuda:0, different from other tensors on cpu (... wrapper_CUDA__index_select) Cause: the load used the `balanced` (device_map) strategy. accelerate's device_map only dispatches components it LOADS — injected pre-loaded components keep whatever device they're on (CPU) with NO offload hook, so a forward pass mixes cuda inputs with cpu weights. Fix: when incremental cached components are injected, force a HOOK-based offload (model → group → sequential), which add offload hooks to EVERY component in the pipeline (injected included). model CPU offload keeps only the active ~7 GB 4-bit expert resident, so it fits easily and is fast. device_map strategies (balanced/disk) are skipped in this case. Safety net: a non-OOM generation failure that looks like a device mismatch now retries ONCE with the incremental cache disabled (plain build = pre-cache behaviour), so generation self-heals even if the injection path is imperfect instead of wedging the queue. (Load-time failures already fall back this way.) Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
A Wan2.2 A14B pipeline that only loads via offload could never be cached — diffusers can't save_pretrained an offloaded pipe, so every load re-quantized from bf16 shards (~14 min) and risked OOM. Cache it per-component instead: - pipeline_cache: per-component subdirs each with their own completion marker + signature (component_dir / component_valid / save_component), atomic via .building + sweep_stale(). A load evicted before all components are cached keeps the finished ones; the next load mixes cached + freshly-built (now-cached) ones, converging to a full cache. (Plus the earlier sweep_stale / _unsavable_reason robustness.) - hf_loading.build_cached_components: for each bnb-quantizable heavy component, load from the per-component cache if present, else build fresh from the model — each ONE AT A TIME on the GPU (where bnb quantizes it to uniform 4-bit), saved, then moved to CPU + empty_cache. Peak VRAM is a single component even when the whole model doesn't fit; uniform 4-bit by construction (no bf16/4-bit mix). Verified the bnb round-trip (load 4-bit on GPU -> save_pretrained -> to('cpu') -> reload). - video._load_video_pipeline: inject the cached components like GGUF components so the existing offload ladder assembles them. Kill-switch CODERAI_INCREMENTAL_CACHE=0. Fully fallback-safe: on any load failure the request handler retries once with use_incremental_cache=False (plain build = today's behaviour). Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
Three fixes for the Wan2.2-VACE-Fun OOM cascade: 1. Offload retry death-spiral (the OOM driver). When from_pretrained OOMs MID-LOAD, `pipe` is never assigned, so the ~20 GB already placed on the GPU is pinned by the EXCEPTION'S TRACEBACK frames (their locals hold the partial model), not by `pipe`. _clear_mem() frees via `pipe` (None) so it reclaims nothing, and the next balanced step recomputes its GPU budget from a card still ~20 GB full (70%->17.5, 60%->3.4, 40%->2.3 GiB — all doomed). Drop `e.__traceback__` and run _clear_mem() OUTSIDE the except block (where sys.exc_info no longer pins the frames) so the stranded tensors are collectable before the next attempt. Applied to the balanced chain and the full-GPU and model-offload handlers. 2. Observability. The load-failure and generation-failure handlers raised HTTPException(500) with the cause only in the HTTP detail — debug.log showed a bare "Response status: 500". Log the full traceback in all three handlers. 3. Pipeline cache robustness. diffusers refuses save_pretrained on a CPU/seq offloaded pipeline, and a device_map/disk-offloaded one has meta tensors — either way the save half-writes a .building dir then a killed process leaves it orphaned (tens of GB). Add _unsavable_reason() to skip those cleanly (no junk), and sweep_stale() to delete orphaned *.building dirs. A full-GPU load stays savable and still caches the 4-bit pipeline. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Two eviction fixes folded into 0.1.30: - VRAM eviction: a diffusers pipeline under device_map/accelerate offload (balanced/disk/model/sequential) or bitsandbytes quantization REJECTS .to('cpu') — the naive move silently left the weights resident and stranded VRAM, feeding the OOM death-spiral where the next load OOMs on a near-full card. Route pipelines (those exposing `components`) through the thorough _free_pipeline_vram (remove accelerate hooks -> drop component refs -> empty_cache); plain non-pipeline models still use a simple .to('cpu'). - RAM eviction: never unload the LIVE model (active_in_vram OR current_model_key) that a request loop keeps reusing. An offloaded pipeline's host RAM IS the live model, so evicting it between same-model requests just forces an immediate reload — churning VRAM/RAM and re-stranding device_map weights while freeing nothing lasting. The last-resort active-model eviction now fires only for a STALE active model (no longer the current one). Co-Authored-By:
-
- 27 Jun, 2026 3 commits
-
-
Stefy Lanza (nextime / spora ) authored
Two bugs let a torch video pipeline run on the nvidia-gguf engine (caps={"gguf"}) when the nvidia engine was briefly down mid-restart: 1. _INFERENCE_PATHS listed "/v1/videos/generations" (plural) but the endpoint is "/v1/video/generations" (singular, video.py:3104). So video requests failed is_inference_path(), skipped all capability-aware routing, and fell through to registry.primary() — which returns the first HEALTHY engine when the primary (nvidia) is unhealthy, i.e. the gguf sibling. Fixed the path. 2. pick_engine() step 5 fell back to a capability-BLIND least_loaded(None) pick, so a typed request could still land on an engine lacking the capability. Now a typed request (transformers/gguf/whisper) only ever picks a capable engine — preferring the primary when it can serve the cap (request queues / caller retries), else 503 — instead of mis-routing to an incompatible engine. Net: video generation (cap=transformers) routes to the nvidia engine only; if it is busy it queues there, if it is restarting the caller retries — it never runs a torch pipeline on the gguf-only engine. Bumps to 0.1.30. Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
The context-free VRAM query used gpu_memory(), which (via NVML/nvidia-smi) reports every physical NVIDIA card regardless of CUDA_VISIBLE_DEVICES. So the Vulkan/Radeon engine (CUDA_VISIBLE_DEVICES="") showed the NVIDIA 3090's VRAM in its status instead of its own AMD card. Switch the per-engine status poll (api/app.py) and the capability probe (broker/capabilities.py) to visible_gpu_memory(), which honours CUDA_VISIBLE_DEVICES (index OR UUID; empty -> no CUDA cards). Add gpu_query.amd_gpu_memory() (amdgpu sysfs, driver-free) as the fallback so the Radeon engine reports its real card; capabilities keeps its torch-loaded guard so the torch-free front still enumerates the whole node. Bumps to 0.1.29. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Reporting VRAM via torch.cuda.mem_get_info lazily creates the CUDA primary context (~256 MiB on an RTX 3090). An engine that never loads a torch model (the GGUF/llama.cpp engine) therefore pinned ~256 MiB just to answer health polls, the capability probe and the load-path eviction check — and that stray context was enough to tip a borderline 4-bit Wan2.2 A14B video load into OOM. New codai/models/gpu_query.py queries VRAM without a context: pynvml first, nvidia-smi fallback, torch only if a context already exists. visible_gpu_memory() scopes to the engine's cards via CUDA_VISIBLE_DEVICES (matched by index OR UUID; empty value -> no CUDA cards, e.g. the Vulkan/Radeon engine). Wired into the idle health poll (api/app.py), the capability probe (broker/capabilities.py) and the load-path free-VRAM checks (models/manager.py: _get_free_vram_gb / _free_vram_snapshot). Adds nvidia-ml-py to requirements-nvidia.txt and the update overlay. Bumps to 0.1.28. Net: the GGUF engine sits at 0 MiB while idle (its real context comes from llama.cpp on model load and is freed on unload), returning the headroom that makes the A14B load fit instead of OOM-ing. Co-Authored-By:
-
- 26 Jun, 2026 7 commits
-
-
Stefy Lanza (nextime / spora ) authored
The diffusers pipeline is cached/reused across image requests, so the adapters _apply_loras added last time linger; re-loading the same fighter then raised "Adapter name <x> already in use", and stale adapters from a different request accumulated. Track the request adapters on the pipeline and delete them (plus a defensive per-name delete) before re-loading, leaving acceleration adapters (__accel__/__accel_2__) untouched. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Loading a trained LoRA at inference (e.g. a fighter LoRA on an image pipeline) crashed with "cannot import name AwqGEMMQuantLinear" because peft dispatches AWQ for any non-bnb target when gptqmodel is installed, and gptqmodel 7.1.0 renamed that class to AwqGEMMLinear. The alias shim existed only in the training path. Extract it to codai/models/peft_compat.ensure_peft_awq_compat() (cached) and call it before load_lora_weights in codai/api/images._apply_loras and the acceleration fuse path; the trainer now delegates to the shared shim. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The 0.1.24 nesting blindly appended the bundled sub-app name to the incoming X-Forwarded-Prefix, which double-counted it when the outer proxy mounts the app under the SAME name and proxies to our /<app>/ path (topology B): outer sends "/township" -> container emitted "/township/township", so every township fetch() 404'd ("Not found" -> JSON parse error in the UI). Per-app maps now only append our name when the incoming prefix doesn't already end in it, so both topologies work: outer-wraps-everything ("/ai" -> "/ai/town ship") and outer-mounts-same-name ("/township" -> "/township"). Direct (no outer proxy) still yields "/township". Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
Internal container nginx is now chain-aware: it prefers the outer proxy's X-Forwarded-Proto/Host/Prefix and nests bundled sub-app prefixes under any outer prefix (outer /ai + /township -> /ai/township). Fixes characters/ environments thumbnails and other absolute/sub-path URLs 404ing when the all-in-one container runs behind a second reverse proxy. Documented the required outer-proxy headers in docs/reverse-proxy-nginx.md. LoRA training loads its base pipeline outside the model manager (and unloads all manager models first), so the engine reported 0 loaded models mid-train. Surface the active training base model via active_training_model() so the engines card reflects the busy GPU. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
- Show which engine each task runs on (badge on active AND history rows; backend already tags t.engine) and an "● processing" badge on the engines card (new inflight/processing fields in /admin/api/engines). - Image-gen advancement showed only "working…": poll() short-circuited to the synthesized task list whenever the primary had any in-flight request, dropping the engine's real step/total. Now it quick-polls the engine first (image/diffusers gen releases the GIL so it answers with real progress) and only falls back to the synthesized list when a GIL-bound text gen times out. - Surface thermal cooldown on the task entry itself: _merge_engine_tasks marks tasks on a cooling engine with cooling/cooling_message (the row already renders it). Co-Authored-By:
-
↔ Stefy Lanza (nextime / spora ) authoredLoRA add_adapter() crashed: peft's dispatch_awq does `from gptqmodel.nn_modules.qlinear.gemm_awq import AwqGEMMQuantLinear`, but gptqmodel 7.1.0 renamed that class to AwqGEMMLinear. peft calls dispatch_awq for ANY non-bnb target when gptqmodel is installed, so the failed import broke every add_adapter (SDXL/Wan/Z-Image), not just AWQ models. Add _ensure_peft_awq_compat(): alias AwqGEMMQuantLinear -> AwqGEMMLinear so peft's import succeeds and falls through to the correct dispatcher. Called before every add_adapter() (sd15/sdxl/dit/wan). Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
QLoRA training reached VAE encode then failed: the ZImagePipeline loads the VAE in bf16, but the image tensor was fed as float32 -> conv2d "Input type (float) and bias type (BFloat16) should be the same". Feed the VAE its own weight dtype. Co-Authored-By:
-
- 25 Jun, 2026 10 commits
-
-
Stefy Lanza (nextime / spora ) authored
QLoRA training crashed in the Qwen3 TEXT ENCODER: the unsloth build quantizes it too, and loading components piecemeal (AutoModel/AutoencoderKL by subfolder) left bitsandbytes' 4-bit quant-state unreconstructed -> "FP4 quantization state not initialized" / AssertionError in Linear4bit.forward. - Load all components via ZImagePipeline.from_pretrained (the proven inference loader, which loads each 4-bit component correctly), then LoRA-train only the transformer. VAE/text-encoder are encoded once then freed (4-bit modules can't .to('cpu'), so drop refs + free). - Disable gradient checkpointing for the 4-bit path: the recompute can desync bnb's quant-state. At 512/batch-1 the 4-bit base + activations fit without it. Co-Authored-By: -
Stefy Lanza (nextime / spora ) authored
Training on the full bf16 Tongyi-MAI/Z-Image-Turbo was extremely slow — a ~10-min download plus a heavy bf16 model that doesn't fit cleanly on 24 GB. Switch _train_dit to QLoRA: load the transformer in 4-bit (frozen, ~4 GB, no CPU offload) and train the LoRA on top. This trains directly on the already-cached 4-bit (e.g. unsloth) build — no redirect to the full model, no download. - Load transformer with diffusers BitsAndBytesConfig (nf4); an already-4-bit checkpoint (embedded quant config, e.g. unsloth) is loaded as-is via fallback. - Enable gradient checkpointing and force the input-embedding (all_x_embedder) output to require grad so QLoRA grads reach the attention LoRA layers; hooks removed at job end. - Drop the quantized-base -> full-model redirect added earlier. LoRA still applies to the quantized model at inference (identical architecture). Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Env/fighter LoRA training for Z-Image misrouted to _train_sdxl (CLIPTokenizer crash: "NoneType cannot be interpreted as an integer"). Cause: _resolve_base_model_path returns the HF id verbatim (e.g. unsloth/Z-Image-Turbo-unsloth-bnb-4bit) when the model has no local `path`, so the isdir(base_path/transformer) DiT check was False and it fell through to the SDXL/SD15 trainer. - Detect Z-Image by NAME (id contains z-image/zimage) in addition to the diffusers-config class name, so an HF-id base routes to _train_dit. - _train_dit: when the base is a pre-quantized (bnb/nf4/4bit) build, train against the full Tongyi-MAI/Z-Image-Turbo instead (a 4-bit checkpoint isn't a clean full-precision LoRA base); the LoRA still applies to the quantized model at inference. Overridable via lora_train_base_model. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
is_available() reported false in the running engine even though gptqmodel + fast kernels import fine in a fresh process. Two causes: 1. capabilities() cached a DEGRADED result (gptqmodel imported but the inner gptqmodel.utils.backend BACKEND import transiently came up empty, e.g. when the first call landed mid model-load). That empty-backends result stuck for the whole process life, so the settings page said "GPTQModel not installed" until restart. Now a degraded (available-but-no-backends) result is NOT cached — re-detect next call; only a clean positive or a genuine ImportError is cached. 2. is_available() gated on a SPECIFIC fast kernel being detected. GPTQModel always has a Triton/torch fallback and picks the kernel at load, so availability now gates only on gptqmodel importing; backends stay informational. Also: /admin/api/quantize-capabilities re-detects live (capabilities(refresh=True)) so the settings page never serves a stale cache. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
1. Model load/unload (and every is_admin-gated front action) returned 401: is_admin probed prim.url + /admin/api/status on the engine, but that route is front-only (removed from the engine in def78c18), so it 404'd → never 200 → Unauthorized. Add an engine-side admin-gated /admin/api/whoami and point is_admin at it. 2. Township reference generation (characters/environments pages — _run_regen_job and _run_create_profile_job) hardcoded 768x512/512x512 + 28 steps, ignoring the Run page. Add _ref_gen_res_steps(args) (honors keyframe_size/keyframe_steps, re-read from the saved config so edits apply without restart) and use it at the reference generators; generate_character/generate_environment now forward `steps` (the server already accepts it). Keyframes/match videos already honored the config. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
coderai's trainer only targeted SDXL/SD1.x U-Nets, producing unet.-prefixed LoRAs that silently no-op on Z-Image's DiT (ZImageTransformer2DModel, transformer. keys, Qwen3 text encoder) — "No LoRA keys associated to ZImageTransformer2DModel found with prefix='transformer'". The lora_train_base_model=SDXL workaround can't make a LoRA that loads on Z-Image. Add _train_dit: a native flow-matching DiT LoRA trainer for Z-Image, reusing the existing job/progress/checkpoint/queue infra (modeled on the Wan video DiT trainer, which is video-only). All deps (ZImageTransformer2DModel, ZImagePipeline, PEFT, Qwen3) are already in the main venv — no ai-toolkit, no separate venv. Reverse-engineered from diffusers pipeline_z_image.py so training matches inference: chat-template + Qwen3 hidden_states[-2] masked per-sample embed list; AutoencoderKL latents scaled (lat-shift)*scale; list-based transformer I/O with normalized timestep (1000-t)/1000; RAW target = x0-noise (Z-Image negates the model output); timesteps sampled from the discrete turbo schedule (set_timesteps mu-shift) to keep the distilled model sharp; save via ZImagePipeline.save_lora_weights (transformer. keys) so _apply_loras' load_lora_weights applies it. _train_lora_sync routes a Z-Image DiT base to _train_dit (detected via model_index/transformer config _class_name); other DiTs (Flux/SD3) still raise with guidance. v1 — needs one validation train; recipe + knobs in docs/zimage-lora-training.md. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
After the GGUF-isolation split, a torch engine and a gguf engine share one NVIDIA card but each can only evict its OWN models. So loading an image model on the torch engine while the gguf engine still holds a resident GGUF text model failed: local eviction "freed 0.0 GB ... VRAM held elsewhere" → CUDA OOM. Fix — co-located VRAM release: - front passes each engine its same-GPU siblings' internal URLs via CODERAI_COSITED_URLS (matched by identical CODERAI_ENGINE_GPUS selectors). - engine registers an external_vram_releaser that POSTs to each sibling's new /internal/evict-vram when local eviction can't free enough. - /internal/evict-vram → manager.release_idle_vram(): evicts all idle (non-busy) models and returns GB freed; busy/actively-serving models are left alone. Symmetric: the gguf engine can likewise reclaim VRAM from the torch engine's idle diffusers models. Driver-level free VRAM (cross-process) is re-checked after each releaser, so the loader proceeds once the sibling has freed enough. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Two fixes after the GGUF process-isolation split: 1. Routing: a model with engine=nvidia in its config (a hard pin) but a GGUF body 503'd, because the torch nvidia engine dropped the `gguf` capability. pick_engine now falls back to the co-located "<name>-gguf" sibling (same card) when the pinned engine can't serve the required capability (and symmetrically for a gguf-pinned model needing transformers). The pin's intent — that physical card — is honoured. 2. Config-save visibility: the cached-models scan is stale-while-revalidate, so the first read after an invalidation returned stale and refreshed in the background — a saved config didn't appear until a second refresh. _invalidate_cache_scan now sets force_sync so the next read recomputes synchronously (fresh immediately). The system worker invalidates on reload-config, so a save shows on the next page load. Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
Co-Authored-By:
-
Stefy Lanza (nextime / spora ) authored
The post-delete redirect used `(window.ROOT_PATH||'') + '/matches'`, but window.ROOT_PATH was never defined, so behind the nginx /township mount it redirected to the bare /matches — which hits the coderai front, not the township tool → 404. The mount shim wraps fetch()/EventSource() to add the prefix but not location.href assignments. Define window.ROOT_PATH=P in the shim so the redirect (and any future ROOT_PATH use) resolves to /township/matches; with no prefix (direct access) it stays unset and the bare /matches path is correct. Co-Authored-By:
-
- 24 Jun, 2026 8 commits
-
-
Stefy Lanza (nextime / spora ) authored
llama.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: -
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.
-