trigger build

computility-run.yaml

@@ -1,48 +1,53 @@
 concurrency: 1
 command:
-    - bash
+    - python3
-    - -c
+    - -m
-    - >-
+    - vllm.entrypoints.openai.api_server
-        python3 /workspace/qwen3_6_scripts/patch_chat_template.py /model 2>&1 || echo '[runtime] chat template patch failed';
+    - --model
-        exec python3 -m vllm.entrypoints.openai.api_server
+    - /model
-        --model /model
+    - --served-model-name
-        --served-model-name llm
+    - llm
-        --max-model-len 131072
+    - --max-model-len
-        --gpu-memory-utilization 0.92
+    - '131072'
-        --trust-remote-code
+    - --gpu-memory-utilization
-        -tp 4
+    - '0.92'
-        --max-num-seqs 2
+    - --trust-remote-code
-        --disable-log-requests
+    - -tp
-        --disable-frontend-multiprocessing
+    - '4'
-        --max-num-batched-tokens 4096
+    - --max-num-seqs
-        --enable-chunked-prefill
+    - '2'
-        --max-seq-len-to-capture 32768
+    - --disable-log-requests
-        --enable-auto-tool-choice
+    - --disable-frontend-multiprocessing
-        --tool-call-parser qwen3_coder
+    - --max-num-batched-tokens
-        --reasoning-parser qwen3
+    - '4096'
-        --enable-prefix-caching
+    - --enable-chunked-prefill
-        --enforce-eager
+    - --max-seq-len-to-capture
-        --dtype half
+    - '32768'
+    - --enable-auto-tool-choice
+    - --tool-call-parser
+    - qwen3_coder
+    - --reasoning-parser
+    - qwen3
+    - --enable-prefix-caching
+    - --enforce-eager
+    - --dtype
+    - half
 env:
     - name: VLLM_ENGINE_ITERATION_TIMEOUT_S
-      value: '3600'
+      value: 3600
     - name: BI100_MAX_NUM_SEQS
-      value: '2'
+      value: 2
-    # --- MoE kernel selection ---
     - name: BI100_MOE_COREX_DIRECT_ROUTED
-      value: '1'
+      value: 1
     - name: BI100_MOE_COREX_TOPK_SOFTMAX
-      value: '1'
+      value: 1
-    # --- GDN kernel selection ---
     - name: BI100_GDN_COREX_PACKED_DECODE
-      value: '1'
+      value: 1
-    # --- Hybrid KV/GDN cache ---
     - name: BI100_HYBRID_KV_ACCOUNTING
       value: full_attention
     - name: BI100_GDN_CACHE_POLICY
       value: admission64
     - name: BI100_GDN_RESTORE_MODE
       value: hybrid64
-    # --- Image fetch timeout (container network) ---
     - name: VLLM_IMAGE_FETCH_TIMEOUT
-      value: '10'
+      value: 10

qwen3_6_scripts/patch_ops.sh

@@ -1,3 +1,5 @@
+echo "[build] trigger 20260901"
+echo "[build] trigger 202609011237"
 #!/usr/bin/env bash
 # BI-V100 patch script for Qwen3.6-35B-A3B (Qwen3_5 MoE architecture)
 #

qwen3_6_scripts/serving_chat.py

@@ -410,9 +410,24 @@ class OpenAIServingChat(OpenAIServing):
 
             if request.max_tokens is not None:
                 request.max_tokens = min(request.max_tokens, _adaptive_cap)
+            else:
+                # [BI100] max_tokens=None means "no limit" in the OpenAI API.
+                # Cap to _adaptive_cap to prevent generation timeout (300s).
+                request.max_tokens = _adaptive_cap
 
             default_max_tokens = min(self.max_model_len - _prompt_len,
                                      _adaptive_cap)
+
+            # [BI100] Qwen3 tool calling fix: greedy decoding (temperature=0)
+            # causes thinking to degrade, so the model never generates
+            # <tool_call> tags.  Force temperature≥0.6 when tools are active.
+            if (request.tools and request.tool_choice in ("auto", None)
+                    and (request.temperature is None
+                         or request.temperature < 0.6)):
+                request.temperature = 0.6
+                if request.top_p is None or request.top_p > 0.95:
+                    request.top_p = 0.95
+
             if request.use_beam_search:
                 sampling_params = request.to_beam_search_params(
                     default_max_tokens)