Disable truss-transfer range downloads: HF CDN now pins signed URLs to the resolving request's Range header, which 403s chunked range reuse

truss/config_nvfp4_4gpu.yaml

@@ -51,6 +51,12 @@ model_cache:
     use_volume: true
 
 environment_variables:
+  # huggingface.co signs its CDN redirects for the exact Range header of the
+  # resolving request, so truss-transfer's range mode (probe bytes=0-0, reuse the
+  # URL for real chunks) gets 403s on every shard. Whole-file streaming avoids the
+  # pinning; 16 parallel files keep the 466GB transfer under ~10 minutes.
+  TRUSS_TRANSFER_USE_RANGE_DOWNLOAD: "0"
+  TRUSS_TRANSFER_NUM_WORKERS: "16"
   # NVFP4 is Blackwell-only and needs the modelopt_fp4 path plus two disable flags
   # (see start_server.sh). GLM5V_CKPT must match volume_folder above.
   GLM5V_CKPT: /app/model_cache/glm5v_nvfp4