feat: dynamically clean incompatible architectures

README.md

@@ -55,6 +55,8 @@ Optional tuning:
 - `MODELHUB_QUEUE_CLEANUP_INTERVAL_CYCLES` default `120`; cleanup also runs once at startup
 - `MODELHUB_QUEUE_CLEANUP_READ_CONCURRENCY` default `6`
 - `MODELHUB_QUEUE_CLEANUP_REPORT_PATH` default `.modelhub_state/queue_cleanup_latest.json`
+- `MODELHUB_ARCHITECTURE_BLACKLIST_PATH` default `.modelhub_state/architecture_compatibility_blacklist.json`
+- `MODELHUB_ARCHITECTURE_BLOCK_TTL_DAYS` default `30`
 - `MODELHUB_RECENT_MODEL_RESERVE_SLOTS` default `10` per account
 - `MODELHUB_DYNAMIC_OLD_MODEL_CLEANUP_RESERVE_SLOTS` default `5` per account
 - `MODELHUB_RECENT_MODEL_DAYS` default `7`
@@ -115,13 +117,14 @@ says that the selected framework does not support the model or architecture, the
 runner learns an exact GPU + framework + task type + architecture block from the
 candidate repository's `config.json`. Repository names are never used as
 architecture evidence. Exact `architectures` values take priority and
-`model_type` is used only when `architectures` is absent; missing metadata does
-not create a block. Generic unsupported operators, attention backends, GPU types,
+`model_type` is also retained when the runtime explicitly says Transformers does
+not recognize that type; missing metadata does not create a block. Generic unsupported operators, attention backends, GPU types,
 quantization failures, and OOMs cannot enter this blacklist. A newer success for
 the same exact combination clears the block, and otherwise it expires after 30
 days. Set `MODELHUB_ARCHITECTURE_BLOCK_TTL_DAYS` to a value from 1 to 365 to
 change that window. The stats report exposes `architectureCompatibilityBlocks`
-and per-GPU/framework block counts.
+and per-GPU/framework block counts. The live snapshot is written to
+`.modelhub_state/architecture_compatibility_blacklist.json`.
 
 Before a candidate reaches the submit queue, failure-informed preflight checks
 the actual ModelScope repository structure and file sizes. Non-GGUF text
@@ -138,7 +141,7 @@ capacities with
 `MODELHUB_GPU_MEMORY_GIB_JSON`, for example
 `{"New_gpu": 64}`.
 
-At poller startup, the same deterministic memory gate is applied to existing
+At poller startup, the same deterministic memory and learned architecture gates are applied to existing
 `waiting` and `running` tasks across every configured account. A task is stopped
 through `PUT /api/async/task/stop-create-contest-task` only when its own current
 recursive repository size, multiplied by ModelHub's observed `1.20` overhead,
@@ -151,6 +154,16 @@ same model or infer failure from historical similarity. The cleanup repeats
 every 120 poll cycles by default and writes its full evidence report to
 `.modelhub_state/queue_cleanup_latest.json`.
 
+Architecture cleanup joins each active task to the locally recorded submission
+or ledger entry to recover its exact framework and task type, then reads the
+model's `config.json`. Only an exact GPU + framework + task type + architecture
+blacklist hit can authorize cancellation. Matching waiting tasks are stopped;
+running tasks remain protected and their state is rechecked again immediately
+before the stop call. Failed outcomes are synchronized every three poll cycles.
+When fixed `MODEL_NOT_SUPPORTED` text adds a new blacklist entry, a lightweight
+architecture-only cleanup runs immediately without repeating repository-size or
+model-age scans.
+
 Each account dynamically reserves its last 10 known-capacity positions for
 models updated within seven days. If an account's discovered limit is 100, 200,
 or 500, older models stop at positions 90, 190, or 490 respectively. Old-model
@@ -270,12 +283,15 @@ Version `2026.08.12.2` learns conservative, expiring GPU/framework/architecture
 compatibility blocks only from explicit ModelHub failure text, matches candidate
 `config.json` metadata instead of repository names, and lets newer success
 evidence clear stale blocks.
+Version `2026.08.12.3` extracts unsupported `model_type`/`architectures` from the
+platform's fixed failure wording, persists a dynamically growing blacklist, and
+immediately removes exact-matching waiting tasks with two active-state checks.
 
 ## Deploy
 
 Create a tag and submit the repository URL plus tag in "我的适配智能体".
 
 ```bash
-git tag agent-v20
-git push origin agent-v20
+git tag agent-v21
+git push origin agent-v21
 ```