upload config data
Browse files- config.json +83 -0
- configscript.sh +7 -0
config.json
ADDED
@@ -0,0 +1,83 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"mlp_bias": false,
|
3 |
+
"attn_bias": false,
|
4 |
+
"rotary_base": 10000,
|
5 |
+
"rotary_scaling": {
|
6 |
+
"beta_fast": 32,
|
7 |
+
"beta_slow": 1,
|
8 |
+
"factor": 40,
|
9 |
+
"mscale": 1.0,
|
10 |
+
"mscale_all_dim": 1.0,
|
11 |
+
"original_max_position_embeddings": 4096,
|
12 |
+
"type": "yarn"
|
13 |
+
},
|
14 |
+
"residual_mlp": false,
|
15 |
+
"disable_weight_only_quant_plugin": false,
|
16 |
+
"moe": {
|
17 |
+
"num_experts": 256,
|
18 |
+
"shared_expert_intermediate_size": 2048,
|
19 |
+
"top_k": 8,
|
20 |
+
"normalization_mode": 3,
|
21 |
+
"sparse_mixer_epsilon": 0.01,
|
22 |
+
"tp_mode": 0,
|
23 |
+
"topk_method": 2,
|
24 |
+
"device_limited_n_group": 8,
|
25 |
+
"device_limited_topk_group": 4,
|
26 |
+
"device_limited_routed_scaling_factor": 2.5
|
27 |
+
},
|
28 |
+
"remove_duplicated_kv_heads": false,
|
29 |
+
"fc_after_embed": false,
|
30 |
+
"use_input_layernorm_in_first_layer": true,
|
31 |
+
"use_last_layernorm": true,
|
32 |
+
"layer_idx_offset": 0,
|
33 |
+
"architecture": "DeepseekV2ForCausalLM",
|
34 |
+
"dtype": "bfloat16",
|
35 |
+
"vocab_size": 129280,
|
36 |
+
"hidden_size": 7168,
|
37 |
+
"num_hidden_layers": 61,
|
38 |
+
"num_attention_heads": 128,
|
39 |
+
"hidden_act": "swiglu",
|
40 |
+
"logits_dtype": "float32",
|
41 |
+
"norm_epsilon": 1e-06,
|
42 |
+
"runtime_defaults": null,
|
43 |
+
"position_embedding_type": "rope_gpt_neox",
|
44 |
+
"num_key_value_heads": 128,
|
45 |
+
"intermediate_size": 18432,
|
46 |
+
"max_position_embeddings": 163840,
|
47 |
+
"mapping": {
|
48 |
+
"world_size": 4,
|
49 |
+
"gpus_per_node": 8,
|
50 |
+
"cp_size": 1,
|
51 |
+
"tp_size": 4,
|
52 |
+
"pp_size": 1,
|
53 |
+
"moe_tp_size": 4,
|
54 |
+
"moe_ep_size": 1
|
55 |
+
},
|
56 |
+
"quantization": {
|
57 |
+
"quant_algo": "W4A16",
|
58 |
+
"kv_cache_quant_algo": null,
|
59 |
+
"group_size": 128,
|
60 |
+
"smoothquant_val": 0.5,
|
61 |
+
"clamp_val": null,
|
62 |
+
"use_meta_recipe": false,
|
63 |
+
"has_zero_point": false,
|
64 |
+
"pre_quant_scale": false,
|
65 |
+
"exclude_modules": null
|
66 |
+
},
|
67 |
+
"use_parallel_embedding": false,
|
68 |
+
"embedding_sharding_dim": 0,
|
69 |
+
"head_size": 56,
|
70 |
+
"qk_layernorm": false,
|
71 |
+
"rotary_embedding_dim": 56,
|
72 |
+
"moe_inter_size": 2048,
|
73 |
+
"kv_lora_rank": 512,
|
74 |
+
"q_lora_rank": 1536,
|
75 |
+
"qk_nope_head_dim": 128,
|
76 |
+
"qk_rope_head_dim": 64,
|
77 |
+
"v_head_dim": 128,
|
78 |
+
"topk_method": "noaux_tc",
|
79 |
+
"first_k_dense_replace": 3,
|
80 |
+
"moe_layer_freq": 1,
|
81 |
+
"coring_func": "sigmoid",
|
82 |
+
"fp8_format": false
|
83 |
+
}
|
configscript.sh
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
python convert_checkpoint.py --model_dir /home/user/hf/deepseek-v3-bf16 \
|
2 |
+
--output_dir /home/user/hf/deepseek-v3-int4 \
|
3 |
+
--dtype bfloat16 \
|
4 |
+
--tp_size 4 \
|
5 |
+
--use_weight_only \
|
6 |
+
--weight_only_precision int4 \
|
7 |
+
--workers 4
|