first commit

dev/online_quant_configs/MiMo-V2-Flash-w4a8/config.json

@@ -0,0 +1,197 @@
+{
+  "architectures": [
+    "MiMoV2FlashForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_mimo_v2_flash.MiMoV2FlashConfig",
+    "AutoModel": "modeling_mimo_v2_flash.MiMoV2FlashModel",
+    "AutoModelForCausalLM": "modeling_mimo_v2_flash.MiMoV2FlashForCausalLM"
+  },
+  "attention_dropout": 0.0,
+  "attention_value_scale": 0.707,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 16384,
+  "max_position_embeddings": 262144,
+  "model_type": "mimo_v2_flash",
+  "num_attention_heads": 64,
+  "head_dim": 192,
+  "num_hidden_layers": 48,
+  "num_key_value_heads": 4,
+  "layernorm_epsilon": 1e-05,
+  "rope_theta": 5000000,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.40.1",
+  "use_cache": true,
+  "vocab_size": 152576,
+  "partial_rotary_factor": 0.334,
+  "sliding_window": 128,
+  "swa_rope_theta": 10000,
+  "attention_bias": false,
+  "v_head_dim": 128,
+  "hybrid_layer_pattern": [
+    0,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    0
+  ],
+  "add_swa_attention_sink_bias": true,
+  "add_full_attention_sink_bias": false,
+  "sliding_window_size": 128,
+  "attention_chunk_size": 128,
+  "moe_layer_freq": [
+    0,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1,
+    1
+  ],
+  "moe_intermediate_size": 2048,
+  "n_routed_experts": 256,
+  "n_shared_experts": null,
+  "num_experts_per_tok": 8,
+  "norm_topk_prob": true,
+  "scoring_func": "sigmoid",
+  "n_group": 1,
+  "topk_group": 1,
+  "topk_method": "noaux_tc",
+  "routed_scaling_factor": null,
+  "swa_num_attention_heads": 64,
+  "swa_num_key_value_heads": 8,
+  "swa_head_dim": 192,
+  "swa_v_head_dim": 128,
+  "quantization_config": {
+    "config_groups": {
+      "group_0": {
+        "input_activations": {
+          "block_structure": null,
+          "dynamic": true,
+          "group_size": null,
+          "num_bits": 8,
+          "observer": "memoryless",
+          "observer_kwargs": {},
+          "strategy": "token",
+          "symmetric": true,
+          "type": "int"
+        },
+        "output_activations": null,
+        "targets": [
+          "Linear"
+        ],
+        "weights": {
+          "block_structure": null,
+          "dynamic": false,
+          "group_size": null,
+          "num_bits": 8,
+          "observer": "minmax",
+          "observer_kwargs": {},
+          "strategy": "channel",
+          "symmetric": true,
+          "type": "int"
+        }
+      }
+    },
+    "format": "int-quantized",
+    "global_compression_ratio": 1.0,
+    "ignore": [
+      "re:lm_head.*",
+      "re:.*norm.*",
+      "re:.*embed_tokens.*",
+      "re:model.layers.\\d+.mlp.gate\\..*",
+      "re:.*bias.*",
+      "re:.*.mtp.*"
+    ],
+    "kv_cache_scheme": null,
+    "quant_method": "compressed-tensors",
+    "quantization_status": "frozen"
+  }
+}