Bootstrap architecture rules from task history

modelhub_submmit_api/failure_log_inspector.py

@@ -32,6 +32,28 @@ MODEL_ARCHITECTURES_NOT_SUPPORTED_PATTERN = re.compile(
     r"(?:are|is)\s+not\s+supported\s+for\s+now",
     re.IGNORECASE,
 )
+SUBMITTED_FRAMEWORK_PATTERN = re.compile(
+    r"\[submit\]\s*framework\s*:\s*(?P<framework>[A-Za-z0-9_.-]{1,80})",
+    re.IGNORECASE,
+)
+TARGET_DOCKER_IMAGE_PATTERN = re.compile(
+    r"\[submit\]\s*docker_image\s*:\s*(?P<image>\S{1,500})",
+    re.IGNORECASE,
+)
+FRAMEWORK_IMAGE_MARKERS = (
+    ("vllm-customized", ("vllm-customized", "vllm_customized")),
+    ("vllm_fix_tokenizer", ("vllm-fix-tokenizer", "vllm_fix_tokenizer")),
+    ("sentence-transformers", ("sentence-transformers", "sentence_transformers")),
+    ("sherpa-onnx", ("sherpa-onnx", "sherpa_onnx")),
+    ("llamacpp", ("llamacpp", "llama-cpp", "llama.cpp")),
+    ("vllm-mlu", ("vllm-mlu", "vllm_mlu")),
+    ("vllm-016", ("vllm-016", "vllm_016")),
+    ("diffusers", ("diffusers",)),
+    ("transformers", ("transformers",)),
+    ("sglang", ("sglang",)),
+    ("funasr", ("funasr",)),
+    ("vllm", ("vllm",)),
+)
 
 
 def fetch_and_classify_failure_log(
@@ -111,6 +133,10 @@ def classify_failure_archive(
         result["failureUnsupportedArchitectures"] = unsupported_architectures
     if unsupported_model_types:
         result["failureUnsupportedModelTypes"] = unsupported_model_types
+    detected_framework, framework_source = _extract_submitted_framework(runtime_log)
+    if detected_framework:
+        result["failureDetectedFramework"] = detected_framework
+        result["failureDetectedFrameworkSource"] = framework_source
     if classification.needs_llm and llm_classifier is not None and llm_classifier.enabled:
         llm_decision = llm_classifier.classify_failure(
             task_context=dict(task_context or {}),
@@ -137,6 +163,22 @@ def classify_failure_archive(
     return result
 
 
+def _extract_submitted_framework(runtime_log: str) -> tuple[str | None, str | None]:
+    """Recover only explicit or unambiguous target-framework evidence."""
+    explicit = SUBMITTED_FRAMEWORK_PATTERN.search(runtime_log)
+    if explicit:
+        return explicit.group("framework").strip(), "submit_framework"
+
+    image_match = TARGET_DOCKER_IMAGE_PATTERN.search(runtime_log)
+    if image_match is None:
+        return None, None
+    target_image = image_match.group("image").strip().casefold()
+    for framework, markers in FRAMEWORK_IMAGE_MARKERS:
+        if any(marker in target_image for marker in markers):
+            return framework, "target_docker_image"
+    return None, None
+
+
 def _extract_error_lines(runtime_log: str) -> list[str]:
     selected: list[str] = []
     for raw_line in runtime_log.splitlines():