Enhance model caching checks and progress tracking for downloads

- Updated caching methods in MLX, PyTorch, and backend to ensure models are fully downloaded before being marked as cached.
- Improved progress tracking to filter out non-download progress and provide accurate feedback during model downloads.
- Enhanced HFProgressTracker to skip non-byte progress bars and ensure meaningful progress reporting.
- Refactored progress initialization to provide immediate feedback while fetching metadata from HuggingFace.
- Added error handling and logging for better debugging during cache checks and download processes.
This commit is contained in:
Jamie Pine
2026-01-30 21:17:01 -08:00
parent d3393fb940
commit 60a03c56a9
5 changed files with 368 additions and 118 deletions
+145 -43
View File
@@ -22,6 +22,7 @@ class HFProgressTracker:
self._file_downloaded = {} # Track downloaded bytes per file
self._current_filename = ""
self._active_tqdms = {} # Track active tqdm instances
self._hf_tqdm_original_update = None # For monkey-patching hf's tqdm
def _create_tracked_tqdm_class(self):
"""Create a tqdm subclass that tracks progress."""
@@ -90,35 +91,66 @@ class HFProgressTracker:
total = getattr(self, "total", 0)
if total and total > 0:
# Determine if we should report this progress
should_report = True
# Always filter out non-byte progress bars (e.g., "Fetching 12 files")
# These cause crazy percentages because they're counting files, not bytes
if self._is_non_byte_progress(filename):
return result
# When model is cached, also filter out generation-related progress
if tracker.filter_non_downloads:
# Filter out non-download progress bars (e.g., "Segment 1/1" during generation)
# Only report progress for actual file downloads from HuggingFace
should_report = self._is_download_progress(filename)
if not self._is_download_progress(filename):
return result
if should_report:
# Update per-file tracking
tracker._file_sizes[filename] = total
tracker._file_downloaded[filename] = current
# Calculate totals across all files
tracker._total_size = sum(tracker._file_sizes.values())
tracker._total_downloaded = sum(tracker._file_downloaded.values())
# Call progress callback
if tracker.progress_callback:
tracker.progress_callback(
tracker._total_downloaded,
tracker._total_size,
filename
)
# Update per-file tracking
tracker._file_sizes[filename] = total
tracker._file_downloaded[filename] = current
# Calculate totals across all files
tracker._total_size = sum(tracker._file_sizes.values())
tracker._total_downloaded = sum(tracker._file_downloaded.values())
# Only report progress once we have a meaningful total (at least 1MB)
# This avoids the "100% at 0MB" issue when small config
# files are counted before the real model files
MIN_TOTAL_BYTES = 1_000_000 # 1MB
if tracker._total_size < MIN_TOTAL_BYTES:
return result
# Call progress callback
if tracker.progress_callback:
tracker.progress_callback(
tracker._total_downloaded,
tracker._total_size,
filename
)
return result
def _is_non_byte_progress(self, filename: str) -> bool:
"""Check if this progress bar should be SKIPPED (returns True to skip).
We want to track byte-based progress bars. This method identifies
progress bars that count files/items instead of bytes, which would
cause crazy percentages if mixed with our byte counting.
Returns:
True = SKIP this bar (it's not byte-based)
False = TRACK this bar (it counts bytes)
"""
if not filename:
return False
filename_lower = filename.lower()
# Skip "Fetching X files" - it counts files (total=12), not bytes
# Don't skip "Downloading (incomplete total...)" - that IS byte-based
skip_patterns = [
'fetching', # "Fetching 12 files" has total=12 files, not bytes
]
return any(pattern in filename_lower for pattern in skip_patterns)
def _is_download_progress(self, filename: str) -> bool:
"""Check if this is a real download progress bar vs internal processing."""
"""Check if this is a real file download progress bar vs internal processing."""
if not filename or filename == "unknown":
return False
@@ -132,7 +164,7 @@ class HFProgressTracker:
filename_lower = filename.lower()
has_extension = any(filename_lower.endswith(ext) for ext in download_extensions)
# Skip internal progress indicators
# Skip generation-related progress indicators
skip_patterns = ['segment', 'processing', 'generating', 'loading']
has_skip_pattern = any(pattern in filename_lower for pattern in skip_patterns)
@@ -177,22 +209,79 @@ class HFProgressTracker:
tqdm_module.auto.tqdm = tracked_tqdm
# Patch in sys.modules to catch already-imported references
# huggingface_hub uses: from tqdm.auto import tqdm as base_tqdm
# So we need to patch both 'tqdm' and 'base_tqdm' attributes
self._patched_modules = {}
tqdm_attr_names = ['tqdm', 'base_tqdm', 'old_tqdm'] # Various names used
patched_count = 0
for module_name in list(sys.modules.keys()):
if "huggingface" in module_name or module_name.startswith("tqdm"):
try:
module = sys.modules[module_name]
if hasattr(module, "tqdm"):
attr = getattr(module, "tqdm")
# Only patch if it's the original tqdm class (not already patched)
if attr is self._original_tqdm_class or (
hasattr(attr, "__name__") and attr.__name__ == "tqdm"
):
self._patched_modules[module_name] = attr
setattr(module, "tqdm", tracked_tqdm)
for attr_name in tqdm_attr_names:
if hasattr(module, attr_name):
attr = getattr(module, attr_name)
# Only patch if it's a tqdm class (not already patched)
is_tqdm_class = (
attr is self._original_tqdm_class or
(self._original_tqdm_auto and attr is self._original_tqdm_auto) or
(hasattr(attr, "__name__") and attr.__name__ == "tqdm" and
hasattr(attr, "update")) # tqdm classes have update method
)
if is_tqdm_class:
key = f"{module_name}.{attr_name}"
self._patched_modules[key] = (module, attr_name, attr)
setattr(module, attr_name, tracked_tqdm)
patched_count += 1
except (AttributeError, TypeError):
pass
# ALSO monkey-patch the update method on huggingface_hub's tqdm class
# This is needed because the class was already defined at import time
self._hf_tqdm_original_update = None
try:
from huggingface_hub.utils import tqdm as hf_tqdm_module
if hasattr(hf_tqdm_module, 'tqdm'):
hf_tqdm_class = hf_tqdm_module.tqdm
self._hf_tqdm_original_update = hf_tqdm_class.update
# Create a wrapper that calls our tracking
tracker = self # Reference to HFProgressTracker instance
def patched_update(tqdm_self, n=1):
result = tracker._hf_tqdm_original_update(tqdm_self, n)
# Track this progress
with tracker._lock:
desc = getattr(tqdm_self, 'desc', '') or ''
current = getattr(tqdm_self, 'n', 0)
total = getattr(tqdm_self, 'total', 0) or 0
# Skip non-byte progress bars
if 'fetching' in desc.lower():
return result
# Skip until we have a meaningful total (at least 1MB)
# This avoids the "100% at 0MB" issue when small config
# files are counted before the real model files
MIN_TOTAL_BYTES = 1_000_000 # 1MB
if total >= MIN_TOTAL_BYTES:
tracker._total_downloaded = current
tracker._total_size = total
if tracker.progress_callback:
tracker.progress_callback(current, total, desc)
return result
hf_tqdm_class.update = patched_update
patched_count += 1
print(f"[HFProgressTracker] Monkey-patched huggingface_hub.utils.tqdm.tqdm.update")
except (ImportError, AttributeError) as e:
print(f"[HFProgressTracker] Could not monkey-patch hf_tqdm: {e}")
print(f"[HFProgressTracker] Patched {patched_count} tqdm references")
yield
except ImportError:
@@ -209,15 +298,24 @@ class HFProgressTracker:
tqdm_module.auto.tqdm = self._original_tqdm_auto
# Restore patched modules
for module_name, original in self._patched_modules.items():
for key, (module, attr_name, original) in self._patched_modules.items():
try:
module = sys.modules.get(module_name)
if module and original:
setattr(module, "tqdm", original)
setattr(module, attr_name, original)
except (AttributeError, TypeError):
pass
self._patched_modules = {}
# Restore hf_tqdm's original update method
if self._hf_tqdm_original_update:
try:
from huggingface_hub.utils import tqdm as hf_tqdm_module
if hasattr(hf_tqdm_module, 'tqdm'):
hf_tqdm_module.tqdm.update = self._hf_tqdm_original_update
except (ImportError, AttributeError):
pass
self._hf_tqdm_original_update = None
except (ImportError, AttributeError):
pass
@@ -225,13 +323,17 @@ class HFProgressTracker:
def create_hf_progress_callback(model_name: str, progress_manager):
"""Create a progress callback for HuggingFace downloads."""
def callback(downloaded: int, total: int, filename: str = ""):
"""Progress callback."""
if total > 0:
progress_manager.update_progress(
model_name=model_name,
current=downloaded,
total=total,
filename=filename or "",
status="downloading",
)
"""Progress callback.
Note: We send updates even when total=0 (unknown) to provide feedback
during the "incomplete total" phase of huggingface_hub downloads.
The frontend handles total=0 gracefully.
"""
progress_manager.update_progress(
model_name=model_name,
current=downloaded,
total=total,
filename=filename or "",
status="downloading",
)
return callback