Switch to Qwen2.5-3B-Instruct and fix model cache bucket name

- Replace Llama-3.1-8B-Instruct with Qwen/Qwen2.5-3B-Instruct
- Rename model cache bucket to aieng-template-cloudrun-vllm
- Bump MAX_NUM_SEQS from 16 → 32 (3B fits more concurrency on L4)
- Update served model name to qwen2.5-3b-instruct

Author: amrit110

.github/workflows/deploy.yml

@@ -33,13 +33,13 @@ env:
   VLLM_SERVICE:       my-project-vllm   # Cloud Run GPU service (vLLM inference)
 
   # ── Model ─────────────────────────────────────────────────────────────────
-  MODEL_ID:           meta-llama/Llama-3.1-8B-Instruct   # HuggingFace model ID
-  MODEL_SERVED_NAME:  my-model          # Name exposed by the vLLM /v1/models API
-  MODEL_CACHE_BUCKET: my-project-model-cache             # GCS bucket for HF weights
+  MODEL_ID:           Qwen/Qwen2.5-3B-Instruct           # HuggingFace model ID
+  MODEL_SERVED_NAME:  qwen2.5-3b-instruct              # Name exposed by the vLLM /v1/models API
+  MODEL_CACHE_BUCKET: aieng-template-cloudrun-vllm       # GCS bucket for HF weights
 
   # ── vLLM inference settings ───────────────────────────────────────────────
   MAX_MODEL_LEN:      8192   # context window (tokens)
-  MAX_NUM_SEQS:       16     # max concurrent sequences (L4 sweet spot for 8B)
+  MAX_NUM_SEQS:       32     # max concurrent sequences (L4 sweet spot for 3B)
   DTYPE:              bfloat16
 
   # ── Demo app settings ─────────────────────────────────────────────────────

Dockerfile.vllm

@@ -32,12 +32,12 @@ EXPOSE 8080
 #
 # For gated models (e.g. Llama 3), pass --hf-token via env var in deploy.yml.
 CMD [ \
-    "meta-llama/Llama-3.1-8B-Instruct", \
-    "--served-model-name", "my-model", \
+    "Qwen/Qwen2.5-3B-Instruct", \
+    "--served-model-name", "qwen2.5-3b-instruct", \
     "--port", "8080", \
     "--host", "0.0.0.0", \
     "--max-model-len", "8192", \
-    "--max-num-seqs", "16", \
+    "--max-num-seqs", "32", \
     "--dtype", "bfloat16", \
     "--enable-chunked-prefill" \
 ]