Fast parallel downloads with hf_transfer

README.md

@@ -8,7 +8,7 @@ sdk_version: 6.8.0
 python_version: '3.12'
 app_file: app.py
 pinned: false
-short_description: Fast video generation with fp8 optimization
+short_description: Fast video generation with bf16 model
 ---
 
 # Element-16 Video Generator

app.py

@@ -1,8 +1,12 @@
 import os
 import subprocess
 import sys
+from concurrent.futures import ThreadPoolExecutor
+
+# Enable fast downloads
+os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
+os.environ["HF_XET_HIGH_PERFORMANCE"] = "1"
 
-# Cache bust: 2024-05-16-v2
 # Disable torch.compile / dynamo before any torch import
 os.environ["TORCH_COMPILE_DISABLE"] = "1"
 os.environ["TORCHDYNAMO_DISABLE"] = "1"
@@ -111,15 +115,33 @@ DISTILL_LORA_REPO = "SulphurAI/Sulphur-2-base"
 LTX_MODEL_REPO = "Lightricks/LTX-2.3"
 GEMMA_REPO = "Lightricks/gemma-3-12b-it-qat-q4_0-unquantized"
 
-# Download model checkpoints
+# Download model checkpoints in parallel for speed
 print("=" * 80)
-print("Downloading Element-16 dev + distill LoRA + Gemma...")
+print("Downloading Element-16 dev + distill LoRA + Gemma (parallel)...")
 print("=" * 80)
 
-checkpoint_path = hf_hub_download(repo_id=CHECKPOINT_REPO, filename="sulphur_dev_bf16.safetensors")
-distilled_lora_path = hf_hub_download(repo_id=DISTILL_LORA_REPO, filename="distill_loras/ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors")
-spatial_upsampler_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-spatial-upscaler-x2-1.0.safetensors")
-gemma_root = snapshot_download(repo_id=GEMMA_REPO)
+def download_checkpoint():
+    return hf_hub_download(repo_id=CHECKPOINT_REPO, filename="sulphur_dev_bf16.safetensors")
+
+def download_lora():
+    return hf_hub_download(repo_id=DISTILL_LORA_REPO, filename="distill_loras/ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors")
+
+def download_upsampler():
+    return hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-spatial-upscaler-x2-1.0.safetensors")
+
+def download_gemma():
+    return snapshot_download(repo_id=GEMMA_REPO)
+
+with ThreadPoolExecutor(max_workers=4) as executor:
+    future_checkpoint = executor.submit(download_checkpoint)
+    future_lora = executor.submit(download_lora)
+    future_upsampler = executor.submit(download_upsampler)
+    future_gemma = executor.submit(download_gemma)
+
+    checkpoint_path = future_checkpoint.result()
+    distilled_lora_path = future_lora.result()
+    spatial_upsampler_path = future_upsampler.result()
+    gemma_root = future_gemma.result()
 
 print(f"Checkpoint: {checkpoint_path}")
 print(f"Distilled LoRA: {distilled_lora_path}")

requirements.txt

@@ -1,6 +1,8 @@
 transformers==4.57.6
 accelerate
 torch==2.8.0
+hf_transfer
+huggingface_hub[hf_xet]
 einops
 scipy
 av