voxtral: HF_TOKEN passthrough + Mistral-format flags (vLLM rejects HF format)

Third voxtral attempt: image pulled clean (3 min, v0.18.0), entrypoint
parsed correctly, vLLM started, but engine init failed two ways:

1. HF rate-limited the irv-ml1 IP (38.120.94.3) during the metadata
   fetch — 429 Too Many Requests from too many large unauthenticated
   pulls today (heretic, 27b, fish-s2, fish-s1-mini, voxtral). Added
   HF_TOKEN env passthrough; user generates a token at
   https://huggingface.co/settings/tokens and sets VOXTRAL_HF_TOKEN
   in .env.

2. Voxtral uses Mistral's native model format (params.json +
   tekken.json tokenizer + consolidated.safetensors single file),
   NOT HF transformers format (config.json + tokenizer.json + sharded
   .safetensors). vLLM errored with "ensure presence of params.json
   for Mistral models." Fix: pass --load-format=mistral
   --tokenizer-mode=mistral --config-format=mistral to vllm serve.
   Confirmed by inspecting the Voxtral-4B-TTS-2603 HF tree:
   25 files, ships params.json + tekken.json + consolidated.safetensors.

Both fixes baked into compose. User needs to drop their HF_TOKEN into
.env once and recreate.

Side note discovered while debugging: fish-s2 s1-mini variant uses
the tiktoken tokenizer format; the wrapper can't load it (errors with
"NoneType has no attribute encode" on warmup). So s1-mini isn't a
drop-in optimization for s2-pro — different code path needed. Fish
back on s2-pro for now.

stacks/voxtral/.env.example

@@ -30,6 +30,14 @@ VOXTRAL_GPU_DEVICES=1
 # alongside other GPU workloads on the same device.
 VOXTRAL_GPU_UTIL=0.85
 
+# ── HuggingFace auth ─────────────────────────────────────────────────
+# HF_TOKEN — required to dodge 429 rate limits on Voxtral download.
+# HF aggressively throttles unauthenticated IPs that pull large repos.
+# Generate one at https://huggingface.co/settings/tokens (a read-only
+# token is sufficient). Without this, the first model download fails
+# mid-stream and vLLM aborts engine init.
+VOXTRAL_HF_TOKEN=
+
 # ── persistent storage on the host ───────────────────────────────────
 # HF cache — first start pulls the Voxtral checkpoint (~8 GB) into
 # this dir. Persistent across container recreates.

stacks/voxtral/compose.yaml

@@ -28,6 +28,10 @@ services:
     environment:
       - NVIDIA_VISIBLE_DEVICES=${VOXTRAL_GPU_DEVICES:-1}
       - HF_HOME=/root/.cache/huggingface
+      # HF_TOKEN required to dodge 429 rate limits on the model
+      # download (HF aggressively throttles unauthenticated IPs that
+      # pull large repos repeatedly). Set in .env — see .env.example.
+      - HF_TOKEN=${VOXTRAL_HF_TOKEN}
       # vLLM serving args — see https://docs.vllm.ai for full list.
       # We override the default model via cmd args below.
     volumes:
@@ -43,6 +47,13 @@ services:
       - "--port=8000"
       - "--dtype=bfloat16"
       - "--gpu-memory-utilization=${VOXTRAL_GPU_UTIL:-0.85}"
+      # Voxtral uses the native Mistral model format (params.json,
+      # tekken.json tokenizer, consolidated.safetensors) — NOT HF
+      # transformers format. vLLM rejects it with "ensure presence
+      # of params.json" unless these three flags are set.
+      - "--load-format=mistral"
+      - "--tokenizer-mode=mistral"
+      - "--config-format=mistral"
     healthcheck:
       # vLLM-Omni exposes /health for liveness + /v1/models for readiness.
       # /health 200 means the server's listening; /v1/models 200 means