Studio: Auto download transport, Xet memory caps, and stall detection on the hub path

studio/backend/hub/schemas/downloads.py

@@ -26,7 +26,14 @@ class DownloadModelRequest(BaseModel):
     )
     use_xet: bool = Field(
         True,
-        description = "Use Xet parallel chunked transport. Default True; set False for HTTP Range-resume.",
+        description = "Legacy transport flag, superseded by transport_mode. Kept so an older "
+        "frontend or a scripted caller keeps working.",
+    )
+    transport_mode: Optional[Literal["auto", "xet", "http"]] = Field(
+        None,
+        description = "Transport preference. 'auto' (the default in the UI) lets the backend pick "
+        "per machine: it knows this host's RAM, its hf_xet build, and whether Xet has "
+        "been failing here. 'xet'/'http' force one. Omitted -> use_xet decides.",
     )
 
 
@@ -93,6 +100,8 @@ class TransportCapability(BaseModel):
 class TransportCapabilities(BaseModel):
     http: TransportCapability
     xet: TransportCapability
+    auto_resolves_to: Literal["xet", "http"] = "xet"
+    auto_reason: Optional[str] = None
 
 
 class TransportStatusResponse(BaseModel):
@@ -126,7 +135,14 @@ class DownloadDatasetRequest(BaseModel):
     repo_id: str = Field(..., description = "HuggingFace dataset repo ID")
     use_xet: bool = Field(
         True,
-        description = "Use Xet parallel chunked transport. Default True; set False for HTTP Range-resume.",
+        description = "Legacy transport flag, superseded by transport_mode. Kept so an older "
+        "frontend or a scripted caller keeps working.",
+    )
+    transport_mode: Optional[Literal["auto", "xet", "http"]] = Field(
+        None,
+        description = "Transport preference. 'auto' (the default in the UI) lets the backend pick "
+        "per machine: it knows this host's RAM, its hf_xet build, and whether Xet has "
+        "been failing here. 'xet'/'http' force one. Omitted -> use_xet decides.",
     )
 
 

studio/backend/hub/services/datasets/downloads.py

@@ -157,8 +157,15 @@ async def download_dataset_response(
     repo_id = await asyncio.to_thread(resolve_cached_repo_id_case, repo_id, repo_type = "dataset")
     key = _download_job_key(repo_id)
 
-    use_xet = download_lifecycle.resolve_effective_use_xet(body.use_xet)
+    # Off the event loop: resolving "auto" can run the Xet reachability probe, and a blackholed
+    # DNS makes that outlast its 3s budget while every other Studio request waits behind it.
+    use_xet, transport_reason = await asyncio.to_thread(
+        download_lifecycle.resolve_requested_use_xet,
+        getattr(body, "transport_mode", None),
+        body.use_xet,
+    )
     transport = download_lifecycle.resolve_transport(use_xet)
+    logger.info("Download transport for %s: %s (%s)", repo_id, transport, transport_reason)
     from utils.hf_cache_settings import get_hf_cache_paths
 
     cache_paths = get_hf_cache_paths()

studio/backend/hub/services/models/downloads.py

@@ -136,8 +136,15 @@ async def download_model_response(
             detail = f"Invalid gguf_variant: {variant!r}",
         )
     key = _download_job_key(repo_id, variant)
-    use_xet = download_lifecycle.resolve_effective_use_xet(body.use_xet)
+    # Off the event loop: resolving "auto" can run the Xet reachability probe, and a blackholed
+    # DNS makes that outlast its 3s budget while every other Studio request waits behind it.
+    use_xet, transport_reason = await asyncio.to_thread(
+        download_lifecycle.resolve_requested_use_xet,
+        getattr(body, "transport_mode", None),
+        body.use_xet,
+    )
     transport = download_lifecycle.resolve_transport(use_xet)
+    logger.info("Download transport for %s: %s (%s)", repo_id, transport, transport_reason)
     from utils.hf_cache_settings import get_hf_cache_paths
 
     cache_paths = get_hf_cache_paths()