sourroy's picture
Add tiny Qwen3-MoE W8A8-INT8 reference model for vLLM CI (router gate left in bf16)
eeebc71 verified
Raw
History Blame Contribute Delete
2.49 kB
{
"architectures": [
"Qwen3MoeForCausalLM"
],
"attention_bias": false,
"attention_dropout": 0.0,
"decoder_sparse_step": 1,
"dtype": "bfloat16",
"head_dim": 128,
"hidden_act": "silu",
"hidden_size": 2048,
"initializer_range": 0.02,
"intermediate_size": 2048,
"max_position_embeddings": 2048,
"mlp_only_layers": [],
"model_type": "qwen3_moe",
"moe_intermediate_size": 1024,
"norm_topk_prob": false,
"num_attention_heads": 16,
"num_experts": 8,
"num_experts_per_tok": 2,
"num_hidden_layers": 2,
"num_key_value_heads": 2,
"output_router_logits": false,
"quantization_config": {
"algo_config": null,
"exclude": [
"model.layers.0.mlp.gate",
"model.layers.1.mlp.gate",
"lm_head"
],
"export": {
"kv_cache_group": [],
"min_kv_scale": 0.0,
"pack_method": "reorder",
"weight_format": "real_quantized",
"weight_merge_groups": null
},
"global_quant_config": {
"bias": null,
"input_tensors": {
"ch_axis": -1,
"dtype": "int8",
"group_size": null,
"is_dynamic": true,
"is_scale_quant": false,
"mx_element_dtype": null,
"observer_cls": "PerChannelMinMaxObserver",
"qscheme": "per_channel",
"round_method": "half_even",
"scale_calculation_mode": null,
"scale_format": null,
"scale_type": "float",
"symmetric": true
},
"output_tensors": null,
"target_device": null,
"weight": {
"ch_axis": 0,
"dtype": "int8",
"group_size": null,
"is_dynamic": false,
"is_scale_quant": false,
"mx_element_dtype": null,
"observer_cls": "PerChannelMinMaxObserver",
"qscheme": "per_channel",
"round_method": "half_even",
"scale_calculation_mode": null,
"scale_format": null,
"scale_type": "float",
"symmetric": true
}
},
"kv_cache_post_rope": false,
"kv_cache_quant_config": {},
"layer_quant_config": {},
"layer_type_quant_config": {},
"quant_method": "quark",
"quant_mode": "eager_mode",
"softmax_quant_spec": null,
"version": "0.11.2"
},
"rms_norm_eps": 1e-06,
"rope_scaling": null,
"rope_theta": 10000.0,
"router_aux_loss_coef": 0.001,
"sliding_window": null,
"tie_word_embeddings": false,
"transformers_version": "4.57.1",
"use_cache": true,
"use_sliding_window": false,
"vocab_size": 151646
}