A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 02:15 UTC
[Bugfix] add support for 'num_attention_groups' in ModelArchConfigConvertorBase for Step3p5 (#39796)
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
realliujiaxu committed
Apr 16, 2026 at 13:48 UTC
7845379230c48c9103ab71682ee05977dfbbdfce
3 files changed
+20
tests/config/base_model_arch_groundtruth.json
+17
@@ -356,6 +356,23 @@
356
"is_multimodal_model": false,
357
"dtype": "torch.float32"
358
},
359
+ "stepfun-ai/Step-3.5-Flash": {
360
+ "architectures": [
361
+ "Step3p5ForCausalLM"
362
+ ],
363
+ "model_type": "step3p5",
364
+ "text_model_type": "step3p5",
365
+ "hidden_size": 4096,
366
+ "total_num_hidden_layers": 45,
367
+ "total_num_attention_heads": 64,
368
+ "head_size": 128,
369
+ "vocab_size": 128896,
370
+ "total_num_kv_heads": 8,
371
+ "num_experts": 288,
372
+ "is_deepseek_mla": false,
373
+ "is_multimodal_model": false,
374
+ "dtype": "torch.bfloat16"
375
+ },
376
"nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16": {
377
"architectures": [
378
"NemotronHForCausalLM"
tests/config/test_model_arch_config.py
+1
@@ -16,6 +16,7 @@ BASE_TRUST_REMOTE_CODE_MODELS = {
16
"nvidia/Llama-3_3-Nemotron-Super-49B-v1",
17
"nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
18
"XiaomiMiMo/MiMo-7B-RL",
19
+ "stepfun-ai/Step-3.5-Flash",
20
# Excluded: Not available online right now
21
# "FreedomIntelligence/openPangu-Ultra-MoE-718B-V1.1",
22
"meituan-longcat/LongCat-Flash-Chat",
vllm/transformers_utils/model_arch_config_convertor.py
+2
@@ -77,6 +77,8 @@ class ModelArchConfigConvertorBase:
77
"num_key_value_heads",
78
# For ChatGLM:
79
"multi_query_group_num",
80
+ # For Step3p5:
81
+ "num_attention_groups",
82
]
83
# For non-grouped-query attention models, the number of KV heads is
84
# equal to the number of attention heads.