Datasets:

LostGentoo
/

hf-inference-endpoint-benchmarks

hf endpoints deploy NAME --repo ORG/MODEL --framework FRAMEWORK \
    --accelerator {gpu|cpu} --instance-type TYPE --instance-size SIZE \
    --vendor {aws|azure|gcp} --region REGION [--namespace NS] [--task TASK] \
    [--min-replica N] [--max-replica N] [--scale-to-zero-timeout MIN] \
    [--scaling-metric {pendingRequests|hardwareUsage}] [--scaling-threshold F] \
    [--token TOKEN]

hf endpoints update NAME [same flags as deploy, all optional, partial update]
hf endpoints describe NAME [--namespace NS] [--token TOKEN]
hf endpoints ls [--namespace NS]
hf endpoints pause NAME
hf endpoints resume NAME
hf endpoints scale-to-zero NAME
hf endpoints delete NAME
hf endpoints catalog ls
hf endpoints catalog deploy --repo ORG/MODEL

Check hf version before trusting any of this. The installed CLI can be badly stale (e.g. 1.5.0 when 1.20+ exists — hf doesn't warn loudly, just nags "a new version is available" on unrelated commands). Update with curl -LsSf https://hf.co/cli/install.sh | bash -. Newer versions add real flags (--custom-image, -e/--env, --secrets, --type) that a stale install's --help won't show — always re-run --help after updating rather than trusting memorized flag lists (including this file's).

--instance-type examples: nvidia-t4, nvidia-l4, nvidia-a10g, nvidia-l40s, nvidia-a100, nvidia-h100, nvidia-h200, nvidia-rtx-pro-6000 (Blackwell — currently the only Blackwell SKU on IE, aws/us-east-2 only, check GET /v2/provider for current availability since regions/types shift), intel-icl, intel-spr. --instance-size is x1/x2/x4/x8 (GPU count) or CPU vCPU tier.

`--framework` does NOT mean "engine" — this is the single biggest CLI trap

--framework example text says 'e.g. vllm', but this is misleading even on the latest CLI (verified against 1.20.1 source). The server's real EndpointFramework enum (confirmed via GET /openapi.json) is only custom | pytorch | llamacpp. Passing --framework vllm gets relayed byte-for-byte to the API and fails with 422 unknown variant "vllm" — the CLI does no translation, at any version checked.

The actual engine selection lives in a separate JSON field, model.image, which the CLI has no flag for at all:

image: {"huggingface": {}} (empty object) — the real "native, auto-select" path. This is what the Python client defaults to when no engine is forced. The backend picks the best native engine (vLLM/TGI/SGLang) for the model itself; you don't get to force a specific one this way, and you won't know which it picked until you check the boot logs.
image: {"vLLM": {tensorParallelSize, dataParallelSize, maxNumSeqs, maxNumBatchedTokens, kvCacheDtype, healthRoute, port, url}} — the structured object the dashboard's "vLLM configuration" panel writes. Confirmed via a real dashboard-created endpoint: url = "vllm/vllm-openai:v0.23.0" (the official upstream vLLM Docker Hub image, version-pinned — not an HF-custom wrapper), healthRoute = /health, port = 8000. So the "native vLLM" support is literally the standard vLLM project image; you can now construct this block by hand via raw API (PUT) since the URL is no longer a mystery — but the version tag will drift over time, so if hand-constructing, check a fresh dashboard-created endpoint's config first to get the current pinned tag rather than trusting this file indefinitely.
image: {"llamacpp": {...}} — the one native engine actually reachable via --framework llamacpp and real CLI flags, since llamacpp is a genuine EndpointFramework enum value.
image: {"custom": {"url": ..., "healthRoute": ..., "port": ...}} — bring your own container (e.g. official vllm/vllm-openai image), reachable via --framework custom --custom-image ... on current CLI, or raw API model.image.custom. This is the fallback if you need to force vLLM specifically and the auto-select path doesn't pick it — but work out the correct entrypoint/mount-path/args on a cheap instance before pointing it at expensive hardware, since none of that is HF-managed anymore once you're in custom.

Practical recipe to get a specific native engine (e.g. vLLM) without guessing the dashboard's URL: deploy with framework: "pytorch", image: {"huggingface": {}}, appropriate task, and inspect the endpoint logs once running to see which engine was actually selected. If it's not what you wanted, that's the signal to move to the custom image path deliberately.

What the CLI cannot do even after updating

No flag sets the structured image.vLLM engine-tuning fields (tensorParallelSize, maxNumSeqs, maxNumBatchedTokens, kvCacheDtype) — those only exist in the OpenAPI schema, reachable only via raw API PUT on an endpoint already using the native vLLM image (which itself you can't select without knowing its url, see above). llama.cpp's own args (-np, -c, --spec-type) similarly need raw API model.args, not any CLI flag — --container-args explicitly requires --custom-image per current CLI help, so it doesn't reach the native llamacpp image either.

Raw REST API

Base: https://api.endpoints.huggingface.cloud/v2/endpoint/{namespace}/{name}

Auth: Authorization: Bearer $TOKEN — the token needs write role / inference.endpoints.write scope. The hf CLI's own login token may not have this — check with hf auth whoami (or curl -H "Authorization: Bearer $TOKEN" https://huggingface.co/api/whoami-v2) before assuming a write will succeed; a scope mismatch fails as HTTP 403 {"error":"Forbidden...missing permissions: inference.endpoints.write"}, not an auth error.

Read current config

TOKEN=$(cat ~/.cache/huggingface/token)
curl -s -H "Authorization: Bearer $TOKEN" \
  "https://api.endpoints.huggingface.cloud/v2/endpoint/$NAMESPACE/$NAME"

Update config — MUST use PUT with a full body, not PATCH

PATCH returns HTTP 405 Method Not Allowed — this API does full-replace semantics only. Always GET the current config first, edit the fields you need, and PUT the whole thing back (a no-op PUT of the unchanged config is a safe way to verify write access).

curl -X PUT \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "compute": {
      "accelerator": "gpu",
      "instanceType": "nvidia-a10g",
      "instanceSize": "x1",
      "scaling": {
        "minReplica": 0, "maxReplica": 1, "scaleToZeroTimeout": 15,
        "metric": "hardwareUsage", "measure": {"hardwareUsage": 80.0}, "threshold": 80.0
      }
    },
    "model": {
      "repository": "unsloth/Qwen3.5-4B-MTP-GGUF",
      "revision": "COMMIT_SHA",
      "task": "image-text-to-text",
      "framework": "llamacpp",
      "image": {
        "llamacpp": {
          "healthRoute": "/health", "port": 80,
          "url": "ghcr.io/ggml-org/llama.cpp:server-cuda",
          "modelPath": "Model-Q8_0.gguf",
          "ctxSize": 0, "nParallel": 10, "threadsHttp": 20, "nGpuLayers": 9999
        }
      },
      "args": ["-ngl","99","-c","81920","-fa","on","-np","10"],
      "env": {"SOME_VAR": "value"},
      "secrets": {"SOME_SECRET": "value"}
    }
  }' \
  "https://api.endpoints.huggingface.cloud/v2/endpoint/$NAMESPACE/$NAME"

Response includes status.state (initializing / updating / running / paused) and status.message (human-readable, e.g. "No replicas are ready yet", "Endpoint is ready"). Poll describe/GET after any write — updates restart the container (brief downtime + cold start).

Health & inference endpoints (once running)

curl -H "Authorization: Bearer $TOKEN" "$ENDPOINT_URL/health"      # {"status":"ok"}
curl -X POST -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"model":"...", "messages":[...], "max_tokens":N}' \
  "$ENDPOINT_URL/v1/chat/completions"                              # OpenAI-compatible

A 503 {"error":"503 Service Unavailable","code":"SERVICE_UNAVAILABLE"} on every path (health, root, v1/models) with no path-specific difference means the HF gateway itself is responding, not a routing problem — almost always a cold start from scale-to-zero (status.state: "initializing", readyReplica: 0) or mid-updating. Check describe/GET state before debugging further.

Custom router (advanced, load-balancing across replicas)

Enable via a top-level customRouter object on create/update (API-only, no CLI/UI support as of this writing):

"customRouter": {
  "tag": "your-org/your-router-image:1.0.0",
  "env": {},
  "port": 3000
}

Reference implementation queued-least-latency exposes Prometheus metrics at /_custom_router/metrics (custom_router_queue_depth, custom_router_backend_ewma_latency_seconds, custom_router_backend_inflight_requests, etc). Useful when replicas have heterogeneous performance or traffic bursts faster than autoscaling reacts. Raise CUSTOM_ROUTER_LATENCY_THRESHOLD above default (3.0s) for LLM/batching workloads where concurrent-per-replica requests increase throughput; leave at default for diffusion/no-batching-benefit workloads.

Autoscaling reference

Scale-up check runs every ~1 minute; scale-down every ~2 minutes, with a 300s stabilization window after scaling down.
hardwareUsage metric: scale up when avg utilization (CPU or GPU) over a 1-min window exceeds threshold (default 80%).
pendingRequests metric: scale up when avg pending (in-flight + queued, no HTTP status yet) requests exceed 1.5/replica over 20s — a leading indicator vs. the lagging hardware metric.
min-replica 0 (scale-to-zero) requires explicit opt-in; default min is 1 (always-on).
Add header X-Scale-Up-Timeout: 600 on a client request to make the proxy hold (rather than 503) while a replica cold-starts, up to N seconds.

CLI Reference & Raw REST API Fallback

hf endpoints command surface

--framework does NOT mean "engine" — this is the single biggest CLI trap