cjw/rocm_vllm_nightly
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

x

Browse Source
This commit is contained in:
SZCJW
2026-03-31 00:02:49 +08:00
parent 234c59b0b4
commit e59ca145be
1 changed files with 7 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
config.json
+7 -7
View File
@@ -34,7 +34,7 @@
"env": {
"VLLM_USE_TRITON_AWQ": "1"
},
"tool_call_parser": "qwen3_xml",
"tool_call_parser": "qwen3_coder",
"reasoning_parser": "qwen3",
"enable_auto_tool_choice": true,
"served_model_name": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
@@ -48,7 +48,7 @@
"max_num_seqs": "8",
"max_tokens": "32768",
"gpu_util": "0.98",
"tool_call_parser": "qwen3_xml",
"tool_call_parser": "qwen3_coder",
"reasoning_parser": "qwen3",
"enable_auto_tool_choice": true,
"served_model_name": "GLM-4.7-Flash-AWQ",
@@ -63,7 +63,7 @@
"max_num_seqs": "4",
"max_tokens": "8192",
"gpu_util": "0.94",
"tool_call_parser": "qwen3_xml",
"tool_call_parser": "qwen3_coder",
"reasoning_parser": "qwen3",
"enable_auto_tool_choice": true,
"served_model_name": "Qwen3.5-27B-FP8",
@@ -72,16 +72,16 @@
"Qwen3.5-35B-A3B-GPTQ-Int4": {
"local_path": "Qwen3.5-35B-A3B-GPTQ-Int4",
"dtype": "float16",
"quantization": "gptq",
"quantization": "moe_wna16",
"model_impl": "transformers",
"ctx": "32768",
"trust_remote": true,
"valid_tp": [2],
"max_num_seqs": "1",
"max_num_seqs": "4",
"max_tokens": "2048",
"gpu_util": "0.90",
"enforce_eager": true,
"tool_call_parser": "qwen3_xml",
"tool_call_parser": "qwen3_coder",
"reasoning_parser": "qwen3",
"enable_auto_tool_choice": true,
"served_model_name": "Qwen3.5-35B-A3B-GPTQ-Int4",
@@ -96,7 +96,7 @@
"max_num_seqs": "8",
"max_tokens": "32768",
"gpu_util": "0.98",
"tool_call_parser": "qwen3_xml",
"tool_call_parser": "qwen3_coder",
"enable_auto_tool_choice": true,
"served_model_name": "Qwen3.5-35B-A3B-FP8",
"hf_model_id": "Qwen/Qwen3.5-35B-A3B-FP8"