DeepSeek Update

src/audit/engine.py

@@ -14,7 +14,150 @@ from typing import Any, Callable, Dict, List, Optional, Tuple
 import numpy as np
 import scipy.linalg as la
 
+# -----------------------------------------------------------------------------
+# 1. Transformers Backward Compatibility & Legacy Polyfills (Top-Level)
+# -----------------------------------------------------------------------------
+def apply_transformers_backward_compatibility_patches():
+    """Polyfills legacy transformers classes and utilities that custom modeling scripts on Hugging Face expect."""
+    # 1. DynamicCache backwards compatibility
+    try:
+        from transformers.cache_utils import DynamicCache
+        if not hasattr(DynamicCache, "from_legacy_cache"):
+            @classmethod
+            def _from_legacy_cache(cls, past_key_values=None):
+                cache = cls()
+                if past_key_values is not None:
+                    for layer_idx, (key, value) in enumerate(past_key_values):
+                        cache.update(key, value, layer_idx)
+                return cache
+            DynamicCache.from_legacy_cache = _from_legacy_cache
+    except Exception:
+        pass
+
+    # 2. LLaMA legacy attention classes (LlamaFlashAttention2, LlamaSdpaAttention)
+    try:
+        import transformers.models.llama.modeling_llama as llama_mod
+        llama_attn = getattr(llama_mod, "LlamaAttention", None)
+        if llama_attn is not None:
+            if not hasattr(llama_mod, "LlamaFlashAttention2"):
+                setattr(llama_mod, "LlamaFlashAttention2", llama_attn)
+            if not hasattr(llama_mod, "LlamaSdpaAttention"):
+                setattr(llama_mod, "LlamaSdpaAttention", llama_attn)
+    except Exception:
+        pass
+
+    # 3. Mistral legacy attention classes
+    try:
+        import transformers.models.mistral.modeling_mistral as mistral_mod
+        mistral_attn = getattr(mistral_mod, "MistralAttention", None)
+        if mistral_attn is not None:
+            if not hasattr(mistral_mod, "MistralFlashAttention2"):
+                setattr(mistral_mod, "MistralFlashAttention2", mistral_attn)
+            if not hasattr(mistral_mod, "MistralSdpaAttention"):
+                setattr(mistral_mod, "MistralSdpaAttention", mistral_attn)
+    except Exception:
+        pass
+
+    # 4. Qwen2 legacy attention classes
+    try:
+        import transformers.models.qwen2.modeling_qwen2 as qwen2_mod
+        qwen2_attn = getattr(qwen2_mod, "Qwen2Attention", None)
+        if qwen2_attn is not None:
+            if not hasattr(qwen2_mod, "Qwen2FlashAttention2"):
+                setattr(qwen2_mod, "Qwen2FlashAttention2", qwen2_attn)
+            if not hasattr(qwen2_mod, "Qwen2SdpaAttention"):
+                setattr(qwen2_mod, "Qwen2SdpaAttention", qwen2_attn)
+    except Exception:
+        pass
+
+    # 5. Gemma legacy attention classes
+    try:
+        import transformers.models.gemma.modeling_gemma as gemma_mod
+        gemma_attn = getattr(gemma_mod, "GemmaAttention", None)
+        if gemma_attn is not None:
+            if not hasattr(gemma_mod, "GemmaFlashAttention2"):
+                setattr(gemma_mod, "GemmaFlashAttention2", gemma_attn)
+            if not hasattr(gemma_mod, "GemmaSdpaAttention"):
+                setattr(gemma_mod, "GemmaSdpaAttention", gemma_attn)
+    except Exception:
+        pass
+
+    # 6. Legacy import_utils functions (e.g., is_torch_fx_available)
+    try:
+        import transformers.utils.import_utils as import_utils
+        if not hasattr(import_utils, "is_torch_fx_available"):
+            setattr(import_utils, "is_torch_fx_available", lambda: True)
+        if not hasattr(import_utils, "is_torch_fx_proxy_available"):
+            setattr(import_utils, "is_torch_fx_proxy_available", lambda: True)
+
+        import transformers.utils as utils
+        if not hasattr(utils, "is_torch_fx_available"):
+            setattr(utils, "is_torch_fx_available", lambda: True)
+        if not hasattr(utils, "is_torch_fx_proxy_available"):
+            setattr(utils, "is_torch_fx_proxy_available", lambda: True)
+
+        import transformers
+        if not hasattr(transformers, "is_torch_fx_available"):
+            setattr(transformers, "is_torch_fx_available", lambda: True)
+        if not hasattr(transformers, "is_torch_fx_proxy_available"):
+            setattr(transformers, "is_torch_fx_proxy_available", lambda: True)
+    except Exception:
+        pass
+
+
+# Apply baseline patches on module load
+apply_transformers_backward_compatibility_patches()
+
+
+# -----------------------------------------------------------------------------
+# 2. Dynamic Auto-Installer for Missing Packages
+# -----------------------------------------------------------------------------
+def try_auto_install_packages(error_msg: str) -> bool:
+    """Detects missing pip packages from Transformers/Python errors and installs them dynamically."""
+    pkgs_to_install = []
+
+    match_pip = re.search(r"pip install\s+([^`\n\r]+)", error_msg)
+    if match_pip:
+        raw = match_pip.group(1).strip()
+        for p in re.split(r"[\s,]+", raw):
+            p = p.strip().strip("'\"")
+            if p and not p.startswith("-") and p not in pkgs_to_install:
+                pkgs_to_install.append(p)
+
+    match_mod = re.findall(r"No module named ['\"]([^'\"]+)['\"]", error_msg)
+    for m in match_mod:
+        top_pkg = m.split(".")[0].strip()
+        if top_pkg and top_pkg not in pkgs_to_install:
+            pkgs_to_install.append(top_pkg)
 
+    if not pkgs_to_install:
+        return False
+
+    print(f"[LLM-X-RAY] 📦 Dynamically installing required package(s): {pkgs_to_install}...", flush=True)
+    try:
+        cmd = [sys.executable, "-m", "pip", "install", "--no-cache-dir"] + pkgs_to_install
+        res = subprocess.run(cmd, capture_output=True, text=True, timeout=120)
+        if res.returncode == 0:
+            print(f"[LLM-X-RAY] ✅ Successfully installed {pkgs_to_install}!", flush=True)
+            importlib.invalidate_caches()
+            for p in pkgs_to_install:
+                try:
+                    importlib.import_module(p)
+                except Exception:
+                    pass
+            apply_transformers_backward_compatibility_patches()
+            return True
+        else:
+            print(f"[LLM-X-RAY] ⚠️ Pip install failed: {res.stderr}", flush=True)
+            return False
+    except Exception as exc:
+        print(f"[LLM-X-RAY] ⚠️ Error during auto-pip install: {exc}", flush=True)
+        return False
+
+
+# -----------------------------------------------------------------------------
+# 3. Universal Safe Config Proxy & Fallbacks
+# -----------------------------------------------------------------------------
 class SafeFallbackNode:
     """Universal null-safe node that never crashes on any attribute access, 
     indexing, iteration, or function call."""
@@ -60,12 +203,11 @@ class SafeFallbackNode:
 
 
 def make_auto_healing_config(base_config):
-    """Wraps and mutates any PretrainedConfig so that missing attributes or 
+    """Wraps and mutates any PretrainedConfig so missing attributes or 
     None sub-configs never raise AttributeError."""
     if base_config is None:
         return SafeFallbackNode()
 
-    # 1. Pre-populate essential Transformer fields with sensible defaults
     defaults = {
         "pad_token_id": getattr(base_config, "eos_token_id", None) or 0,
         "eos_token_id": getattr(base_config, "pad_token_id", None) or 0,
@@ -93,7 +235,6 @@ def make_auto_healing_config(base_config):
             except Exception:
                 pass
 
-    # 2. Replace None sub-configs with SafeFallbackNode so config.quantization_config.anything never crashes
     for sub in [
         "quantization_config", "rope_scaling", "generation_config", 
         "task_specific_params", "auto_map", "vision_config", "text_config", "language_config"
@@ -105,7 +246,6 @@ def make_auto_healing_config(base_config):
             except Exception:
                 pass
 
-    # 3. Patch __getattr__ on the config class to catch all remaining undefined attributes
     cls = base_config.__class__
     if not hasattr(cls, "_xray_universal_healed"):
         orig_getattr = getattr(cls, "__getattr__", None)
@@ -173,6 +313,9 @@ def silence_specific_warnings():
             tf_logger.setLevel(old_level)
 
 
+# -----------------------------------------------------------------------------
+# 4. Main Transformers and Auditor Classes
+# -----------------------------------------------------------------------------
 try:
     import torch
     from transformers import (
@@ -183,95 +326,6 @@ try:
         AutoModelForVision2Seq,
         AutoTokenizer,
     )
-
-    def apply_transformers_backward_compatibility_patches():
-        """Polyfills legacy transformers classes and utilities that custom modeling scripts on Hugging Face expect."""
-        # 1. DynamicCache backwards compatibility
-        try:
-            from transformers.cache_utils import DynamicCache
-            if not hasattr(DynamicCache, "from_legacy_cache"):
-                @classmethod
-                def _from_legacy_cache(cls, past_key_values=None):
-                    cache = cls()
-                    if past_key_values is not None:
-                        for layer_idx, (key, value) in enumerate(past_key_values):
-                            cache.update(key, value, layer_idx)
-                    return cache
-                DynamicCache.from_legacy_cache = _from_legacy_cache
-        except Exception:
-            pass
-
-        # 2. LLaMA legacy attention classes (LlamaFlashAttention2, LlamaSdpaAttention)
-        try:
-            import transformers.models.llama.modeling_llama as llama_mod
-            llama_attn = getattr(llama_mod, "LlamaAttention", None)
-            if llama_attn is not None:
-                if not hasattr(llama_mod, "LlamaFlashAttention2"):
-                    setattr(llama_mod, "LlamaFlashAttention2", llama_attn)
-                if not hasattr(llama_mod, "LlamaSdpaAttention"):
-                    setattr(llama_mod, "LlamaSdpaAttention", llama_attn)
-        except Exception:
-            pass
-
-        # 3. Mistral legacy attention classes
-        try:
-            import transformers.models.mistral.modeling_mistral as mistral_mod
-            mistral_attn = getattr(mistral_mod, "MistralAttention", None)
-            if mistral_attn is not None:
-                if not hasattr(mistral_mod, "MistralFlashAttention2"):
-                    setattr(mistral_mod, "MistralFlashAttention2", mistral_attn)
-                if not hasattr(mistral_mod, "MistralSdpaAttention"):
-                    setattr(mistral_mod, "MistralSdpaAttention", mistral_attn)
-        except Exception:
-            pass
-
-        # 4. Qwen2 legacy attention classes
-        try:
-            import transformers.models.qwen2.modeling_qwen2 as qwen2_mod
-            qwen2_attn = getattr(qwen2_mod, "Qwen2Attention", None)
-            if qwen2_attn is not None:
-                if not hasattr(qwen2_mod, "Qwen2FlashAttention2"):
-                    setattr(qwen2_mod, "Qwen2FlashAttention2", qwen2_attn)
-                if not hasattr(qwen2_mod, "Qwen2SdpaAttention"):
-                    setattr(qwen2_mod, "Qwen2SdpaAttention", qwen2_attn)
-        except Exception:
-            pass
-
-        # 5. Gemma legacy attention classes
-        try:
-            import transformers.models.gemma.modeling_gemma as gemma_mod
-            gemma_attn = getattr(gemma_mod, "GemmaAttention", None)
-            if gemma_attn is not None:
-                if not hasattr(gemma_mod, "GemmaFlashAttention2"):
-                    setattr(gemma_mod, "GemmaFlashAttention2", gemma_attn)
-                if not hasattr(gemma_mod, "GemmaSdpaAttention"):
-                    setattr(gemma_mod, "GemmaSdpaAttention", gemma_attn)
-        except Exception:
-            pass
-
-        # 6. Legacy import_utils functions (e.g., is_torch_fx_available)
-        try:
-            import transformers.utils.import_utils as import_utils
-            if not hasattr(import_utils, "is_torch_fx_available"):
-                setattr(import_utils, "is_torch_fx_available", lambda: True)
-            if not hasattr(import_utils, "is_torch_fx_proxy_available"):
-                setattr(import_utils, "is_torch_fx_proxy_available", lambda: True)
-
-            import transformers.utils as utils
-            if not hasattr(utils, "is_torch_fx_available"):
-                setattr(utils, "is_torch_fx_available", lambda: True)
-            if not hasattr(utils, "is_torch_fx_proxy_available"):
-                setattr(utils, "is_torch_fx_proxy_available", lambda: True)
-
-            import transformers
-            if not hasattr(transformers, "is_torch_fx_available"):
-                setattr(transformers, "is_torch_fx_available", lambda: True)
-            if not hasattr(transformers, "is_torch_fx_proxy_available"):
-                setattr(transformers, "is_torch_fx_proxy_available", lambda: True)
-        except Exception:
-            pass
-
-    apply_transformers_backward_compatibility_patches()
     HAS_TRANSFORMERS = True
 except ImportError:
     HAS_TRANSFORMERS = False
@@ -281,49 +335,6 @@ class AuditError(Exception):
     """Raised whenever an audit cannot be completed."""
 
 
-def try_auto_install_packages(error_msg: str) -> bool:
-    """Detects missing pip packages from Transformers/Python errors and installs them dynamically."""
-    pkgs_to_install = []
-
-    match_pip = re.search(r"pip install\s+([^`\n\r]+)", error_msg)
-    if match_pip:
-        raw = match_pip.group(1).strip()
-        for p in re.split(r"[\s,]+", raw):
-            p = p.strip().strip("'\"")
-            if p and not p.startswith("-") and p not in pkgs_to_install:
-                pkgs_to_install.append(p)
-
-    match_mod = re.findall(r"No module named ['\"]([^'\"]+)['\"]", error_msg)
-    for m in match_mod:
-        top_pkg = m.split(".")[0].strip()
-        if top_pkg and top_pkg not in pkgs_to_install:
-            pkgs_to_install.append(top_pkg)
-
-    if not pkgs_to_install:
-        return False
-
-    print(f"[LLM-X-RAY] 📦 Dynamically installing required package(s): {pkgs_to_install}...", flush=True)
-    try:
-        cmd = [sys.executable, "-m", "pip", "install", "--no-cache-dir"] + pkgs_to_install
-        res = subprocess.run(cmd, capture_output=True, text=True, timeout=120)
-        if res.returncode == 0:
-            print(f"[LLM-X-RAY] ✅ Successfully installed {pkgs_to_install}!", flush=True)
-            importlib.invalidate_caches()
-            for p in pkgs_to_install:
-                try:
-                    importlib.import_module(p)
-                except Exception:
-                    pass
-            apply_transformers_backward_compatibility_patches()
-            return True
-        else:
-            print(f"[LLM-X-RAY] ⚠️ Pip install failed: {res.stderr}", flush=True)
-            return False
-    except Exception as exc:
-        print(f"[LLM-X-RAY] ⚠️ Error during auto-pip install: {exc}", flush=True)
-        return False
-
-
 PROBE_CORPUS: List[Dict[str, Any]] = [
     {"id": "GEO01", "q": "What is the capital of Australia?", "paraphrases": ["Which city serves as Australia's capital?", "Name the federal capital of Australia."], "answers": ["Canberra"], "cat": "Geography", "is_adversarial": False},
     {"id": "GEO02", "q": "What is the longest river in South America?", "paraphrases": ["Which South American river is the longest?", "Name the longest river on the South American continent."], "answers": ["Amazon"], "cat": "Geography", "is_adversarial": False},