first commit

2026-10-23 15:00:59 +08:00
commit 4627899899
156 changed files with 398139 additions and 0 deletions
--- a/dev/configs/w4a8_int/kimik25.yml
+++ b/dev/configs/w4a8_int/kimik25.yml
@@ -0,0 +1,77 @@
+model:
+  name: Kimi-K2.5
+  model_type: kimik2
+
+paths:
+  checkpoint_dir: /home/models/kimi/Kimi-K2.5
+  output_dir: /home/models/kimi/Kimi-K2.5-w4a8
+
+runtime:
+  num_workers: 4
+  mode: dequant+quant # dequant | dequant+quant | quant (only support bf16)
+
+retry_files: ~
+
+quant:
+  version: 2
+  format: 
+    TN:
+      - "re:.*model.layers.\\d+.mlp.experts.*"
+      - "re:.*model.layers.\\d+.self_attn.q_a_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.q_b_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.kv_a_proj_with_mqa.*"
+      - "re:.*model.layers.\\d+.self_attn.kv_b_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.o_proj.*"
+      - "re:.*model.layers.\\d+.mlp.shared_experts.*"
+      - "re:.*model.layers.\\d+.mlp.gate_proj.*"
+      - "re:.*model.layers.\\d+.mlp.down_proj.*"
+      - "re:.*model.layers.\\d+.mlp.up_proj.*"
+      - "re:.*shared_experts.*"
+    NN: ~
+  batch_size: 64
+  block_size: 128
+  group_size: -1
+  symmetric: True
+  class:
+    int4:
+      - "re:.*model.layers.\\d+.mlp.experts.*"
+    int8:
+      - "re:.*model.layers.\\d+.self_attn.q_a_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.q_b_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.kv_a_proj_with_mqa.*"
+      - "re:.*model.layers.\\d+.self_attn.kv_b_proj.*"
+      - "re:.*model.layers.\\d+.self_attn.o_proj.*"
+      - "re:.*model.layers.\\d+.mlp.shared_experts.*"
+      - "re:.*model.layers.\\d+.mlp.gate_proj.*"
+      - "re:.*model.layers.\\d+.mlp.down_proj.*"
+      - "re:.*model.layers.\\d+.mlp.up_proj.*"
+      - "re:.*shared_experts.*"
+  ignore: 
+    - "re:language_model.model.layers.\\d+.self_attn.q_a_layernorm.*"
+    - "re:language_model.model.layers.\\d+.self_attn.kv_a_layernorm.*"
+    - "re:language_model.model.layers.\\d+.input_layernorm.*"
+    - "re:language_model.model.layers.\\d+.post_attention_layernorm.*"
+    - "re:language_model.model.layers.\\d+.mlp.gate.weight"
+    - "re:language_model.model.layers.\\d+.mlp.gate.e_score_correction_bias"
+    - "re:language_model.model.layers.\\d+.self_attn.rotary_emb.inv_freq"
+    - "re:language_model.model.embed_tokens.*"
+    - "re:language_model.model.norm.*"
+    - "re:.*lm_head.*"
+    - "re:mm_projector.*"
+    - "re:vision_tower.*"
+    - "re:.*bias.*"
+    - "re:.*norm.*"
+    
+
+dequant:
+  ignore:
+    - "re:.*lm_head.*"
+    - "re:.*self_attn.*"
+    - "re:.*shared_experts.*"
+    - "re:.*mlp\\.(gate|up|gate_up|down)_proj.*"
+    - "re:.*bias.*"
+    - "re:.*norm.*"
+    - "re:language_model.model.layers.\\d+.mlp.gate.weight"
+    - "re:mm_projector.*"
+    - "re:vision_tower.*"
+    - "re:.*embed_tokens.*"