Add reproducible DGX Spark deployment for Qwen3.8 Flash Next NVFP4

2026-09-17 16:45:50 +08:00
commit 5f3030260e
27 changed files with 1705 additions and 0 deletions
@@ -0,0 +1,11 @@
+# Copy to .env on the Spark; use absolute paths. Do not commit .env.
+HF_CACHE_DIR=/path/to/huggingface
+API_KEY_FILE=/path/to/qwen.key
+API_PORT=8000
+BIND_ADDRESS=0.0.0.0
+MODEL_HTTP_PROXY=
+MODEL_HTTPS_PROXY=
+GPU_MEMORY_UTILIZATION=0.80
+MAX_NUM_SEQS=4
+# Optional offline build: local tag that resolves to the pinned base digest.
+# BASE_IMAGE=local/vllm-base:0bfc7a15