Sync from GitHub 1e79046
Browse files- Dockerfile +2 -1
- main.py +10 -2
Dockerfile
CHANGED
|
@@ -23,6 +23,7 @@ ARG BACKEND=cpu
|
|
| 23 |
# - CUDA cu124 index publishes up to 2.6.0 (auto-resolves to +cu124)
|
| 24 |
# - ROCm 6.2 index publishes up to 2.5.1+rocm6.2 (must include local tag)
|
| 25 |
ARG TORCH_VER_CPU=2.9.0
|
|
|
|
| 26 |
ARG TORCH_VER_NVIDIA=2.6.0
|
| 27 |
ARG TORCH_VER_AMD=2.5.1+rocm6.2
|
| 28 |
|
|
@@ -35,7 +36,7 @@ ENV PIP_NO_CACHE_DIR=1
|
|
| 35 |
|
| 36 |
# Install appropriate PyTorch for the selected backend, then the rest
|
| 37 |
RUN if [ "$BACKEND" = "cpu" ]; then \
|
| 38 |
-
pip install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${TORCH_VER_CPU}; \
|
| 39 |
elif [ "$BACKEND" = "nvidia" ]; then \
|
| 40 |
pip install --no-cache-dir --index-url https://download.pytorch.org/whl/cu124 torch==${TORCH_VER_NVIDIA}; \
|
| 41 |
elif [ "$BACKEND" = "amd" ]; then \
|
|
|
|
| 23 |
# - CUDA cu124 index publishes up to 2.6.0 (auto-resolves to +cu124)
|
| 24 |
# - ROCm 6.2 index publishes up to 2.5.1+rocm6.2 (must include local tag)
|
| 25 |
ARG TORCH_VER_CPU=2.9.0
|
| 26 |
+
ARG TORCHVISION_VER_CPU=0.24.0
|
| 27 |
ARG TORCH_VER_NVIDIA=2.6.0
|
| 28 |
ARG TORCH_VER_AMD=2.5.1+rocm6.2
|
| 29 |
|
|
|
|
| 36 |
|
| 37 |
# Install appropriate PyTorch for the selected backend, then the rest
|
| 38 |
RUN if [ "$BACKEND" = "cpu" ]; then \
|
| 39 |
+
pip install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==${TORCH_VER_CPU} torchvision==${TORCHVISION_VER_CPU}; \
|
| 40 |
elif [ "$BACKEND" = "nvidia" ]; then \
|
| 41 |
pip install --no-cache-dir --index-url https://download.pytorch.org/whl/cu124 torch==${TORCH_VER_NVIDIA}; \
|
| 42 |
elif [ "$BACKEND" = "amd" ]; then \
|
main.py
CHANGED
|
@@ -105,8 +105,16 @@ def prefetch_model_assets(repo_id: str, token: Optional[str]) -> Optional[str]:
|
|
| 105 |
- If CLI is unavailable, falls back to verbose API prefetch.
|
| 106 |
"""
|
| 107 |
try:
|
| 108 |
-
# Enable accelerated transfer
|
| 109 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 110 |
os.environ.setdefault("HF_HUB_ENABLE_XET", "1")
|
| 111 |
|
| 112 |
cache_dir = os.getenv("HF_HOME") or os.getenv("TRANSFORMERS_CACHE") or ""
|
|
|
|
| 105 |
- If CLI is unavailable, falls back to verbose API prefetch.
|
| 106 |
"""
|
| 107 |
try:
|
| 108 |
+
# Enable accelerated transfer only if hf_transfer is installed; otherwise disable to avoid runtime errors on Spaces
|
| 109 |
+
try:
|
| 110 |
+
import importlib.util as _imputil
|
| 111 |
+
if _imputil.find_spec("hf_transfer") is not None:
|
| 112 |
+
os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
|
| 113 |
+
else:
|
| 114 |
+
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
|
| 115 |
+
except Exception:
|
| 116 |
+
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "0"
|
| 117 |
+
# XET acceleration if available; harmless if missing
|
| 118 |
os.environ.setdefault("HF_HUB_ENABLE_XET", "1")
|
| 119 |
|
| 120 |
cache_dir = os.getenv("HF_HOME") or os.getenv("TRANSFORMERS_CACHE") or ""
|