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:
chenjw28
2026-04-17 09:38:30 +08:00
parent d87daadc2f
commit 80a549bdcd
1 changed files with 88 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files
config.json
+88 -30
View File
@@ -20,40 +20,70 @@
"revision": "", "revision": "",
"models": { "models": {
"default": "Qwen3.5-35B-A3B-GPTQ-Int4", "default": "Qwen3.5-35B-A3B-GPTQ-Int4",
"selected": "Qwen3.5-27B-FP8", "selected": "Qwen3.6-35B-A3B-FP8",
"profiles": { "profiles": {
"GLM-4.7-Flash-AWQ": { "Qwen3.6-35B-A3B": {
"local_path": "GLM-4.7-Flash-AWQ", "local_path": "Qwen3.6-35B-A3B",
"ctx": "32768", "dtype": "float16",
"quantization": "auto",
"ctx": "131072",
"max_tokens": "8192",
"trust_remote": true, "trust_remote": true,
"valid_tp": [1, 2], "enforce_eager": false,
"max_num_seqs": "16", "valid_tp": [
"max_tokens": "32768", 2
"gpu_util": "0.98", ],
"max_num_seqs": "12",
"gpu_util": "0.90",
"tool_call_parser": "qwen3_coder", "tool_call_parser": "qwen3_coder",
"reasoning_parser": "qwen3", "reasoning_parser": "qwen3",
"enable_auto_tool_choice": true, "enable_auto_tool_choice": true,
"served_model_name": "GLM-4.7-Flash-AWQ", "served_model_name": "Qwen3.6-35B-A3B",
"hf_model_id": "THUDM/GLM-4.7-Flash-AWQ" "hf_model_id": "Qwen/Qwen3.6-35B-A3B",
"env": {
"VLLM_RPC_TIMEOUT": "300"
}
}, },
"Qwen3.5-27B-FP8": { "Qwen3.6-35B-A3B-FP8": {
"local_path": "Qwen3.5-27B-FP8", "local_path": "Qwen3.6-35B-A3B-FP8",
"dtype": "auto", "dtype": "float16",
"quantization": "auto",
"ctx": "131072", "ctx": "131072",
"max_tokens": "8192",
"trust_remote": true, "trust_remote": true,
"valid_tp": [1, 2],
"max_num_seqs": 32,
"max_tokens": "16384",
"enforce_eager": false, "enforce_eager": false,
"valid_tp": [
2
],
"max_num_seqs": "12",
"gpu_util": "0.90", "gpu_util": "0.90",
"tool_call_parser": "qwen3_coder",
"reasoning_parser": "qwen3",
"enable_auto_tool_choice": true,
"served_model_name": "Qwen3.6-35B-A3B-FP8",
"hf_model_id": "Qwen/Qwen3.6-35B-A3B-FP8"
},
"Qwen3.5-35B-A3B-FP8": {
"local_path": "Qwen3.5-35B-A3B-FP8",
"dtype": "float16",
"ctx": "131072",
"enforce_eager": false,
"trust_remote": true,
"valid_tp": [
1,
2
],
"max_num_seqs": "8",
"max_tokens": "16384",
"gpu_util": "0.94",
"kv_cache_dtype": "fp8", "kv_cache_dtype": "fp8",
"enable_prefix_caching": true, "enable_prefix_caching": true,
"max_num_batched_tokens": 16384, "max_num_batched_tokens": 16384,
"tool_call_parser": "qwen3_coder", "tool_call_parser": "qwen3_coder",
"reasoning_parser": "qwen3", "reasoning_parser": "qwen3",
"enable_auto_tool_choice": true, "enable_auto_tool_choice": true,
"served_model_name": "Qwen3.5-27B-FP8", "served_model_name": "Qwen3.5-35B-A3B-FP8",
"hf_model_id": "RedHatAI/Qwen3.5-27B-FP8-dynamic" "hf_model_id": "Qwen/Qwen3.5-35B-A3B-FP8"
}, },
"Qwen3.5-35B-A3B-GPTQ-Int4": { "Qwen3.5-35B-A3B-GPTQ-Int4": {
"local_path": "Qwen3.5-35B-A3B-GPTQ-Int4", "local_path": "Qwen3.5-35B-A3B-GPTQ-Int4",
@@ -61,7 +91,9 @@
"quantization": "gptq", "quantization": "gptq",
"ctx": "131072", "ctx": "131072",
"trust_remote": true, "trust_remote": true,
"valid_tp": [2], "valid_tp": [
2
],
"max_num_seqs": "6", "max_num_seqs": "6",
"max_tokens": "4096", "max_tokens": "4096",
"gpu_util": "0.90", "gpu_util": "0.90",
@@ -75,31 +107,37 @@
"VLLM_RPC_TIMEOUT": "300" "VLLM_RPC_TIMEOUT": "300"
} }
}, },
"Qwen3.5-35B-A3B-FP8": { "Qwen3.5-27B-FP8": {
"local_path": "Qwen3.5-35B-A3B-FP8", "local_path": "Qwen3.5-27B-FP8",
"dtype": "float16", "dtype": "auto",
"ctx": "131072", "ctx": "131072",
"enforce_eager": false,
"trust_remote": true, "trust_remote": true,
"valid_tp": [1, 2], "valid_tp": [
"max_num_seqs": "8", 1,
2
],
"max_num_seqs": 32,
"max_tokens": "16384", "max_tokens": "16384",
"gpu_util": "0.94", "enforce_eager": false,
"gpu_util": "0.90",
"kv_cache_dtype": "fp8", "kv_cache_dtype": "fp8",
"enable_prefix_caching": true, "enable_prefix_caching": true,
"max_num_batched_tokens": 16384, "max_num_batched_tokens": 16384,
"tool_call_parser": "qwen3_coder", "tool_call_parser": "qwen3_coder",
"reasoning_parser": "qwen3", "reasoning_parser": "qwen3",
"enable_auto_tool_choice": true, "enable_auto_tool_choice": true,
"served_model_name": "Qwen3.5-35B-A3B-FP8", "served_model_name": "Qwen3.5-27B-FP8",
"hf_model_id": "Qwen/Qwen3.5-35B-A3B-FP8" "hf_model_id": "RedHatAI/Qwen3.5-27B-FP8-dynamic"
}, },
"Qwen3-Next-80B-A3B-Instruct-AWQ-4bit": { "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit": {
"local_path": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit", "local_path": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
"dtype": "auto", "dtype": "auto",
"ctx": "131072", "ctx": "131072",
"trust_remote": true, "trust_remote": true,
"valid_tp": [1, 2], "valid_tp": [
1,
2
],
"max_num_seqs": "8", "max_num_seqs": "8",
"max_tokens": "65536", "max_tokens": "65536",
"max_num_batched_tokens": 16384, "max_num_batched_tokens": 16384,
@@ -112,12 +150,32 @@
"served_model_name": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit", "served_model_name": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
"hf_model_id": "cpatonn/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit" "hf_model_id": "cpatonn/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit"
}, },
"GLM-4.7-Flash-AWQ": {
"local_path": "GLM-4.7-Flash-AWQ",
"ctx": "32768",
"trust_remote": true,
"valid_tp": [
1,
2
],
"max_num_seqs": "16",
"max_tokens": "32768",
"gpu_util": "0.98",
"tool_call_parser": "qwen3_coder",
"reasoning_parser": "qwen3",
"enable_auto_tool_choice": true,
"served_model_name": "GLM-4.7-Flash-AWQ",
"hf_model_id": "THUDM/GLM-4.7-Flash-AWQ"
},
"Gemma-4-26B-A4B": { "Gemma-4-26B-A4B": {
"local_path": "gemma-4-26B-A4B", "local_path": "gemma-4-26B-A4B",
"dtype": "auto", "dtype": "auto",
"ctx": "65536", "ctx": "65536",
"trust_remote": true, "trust_remote": true,
"valid_tp": [1, 2], "valid_tp": [
1,
2
],
"max_num_seqs": 16, "max_num_seqs": 16,
"max_tokens": "8192", "max_tokens": "8192",
"enforce_eager": false, "enforce_eager": false,