WpyQwq
643e22ecb9
- 引入 MemoryRouterXL 与 v5/v6 流式多线程训练/编码管线 - 修复 prepare_memory_router_dataset 候选池重建缺陷（mega 家族 3568x 加速，输出逐字节相同） - 修复 v5 被破坏的拒答与多跳标签（train 未知样本 319 -> 16319，multi_hop 平均正例 1.00 -> 2.00） - 同存储预算下 V2-128 v6 逐轴 22/22 通过：Top-1 41.12% -> 94.62%，未知拒答 0.00% -> 100.00% - 记录三条被实测推翻的显然优化（logits_to_keep=1 反而慢 55%、XL 容量未带来收益） - 记忆手术跨架构可移植性 14/14，读写关闭时与原生模型逐位相同
43 lines
1.3 KiB
JSON
43 lines
1.3 KiB
JSON
[
|
|
{
|
|
"model": "H:\\modelscope_cache\\ibm-granite\\granite-3___0-3b-a800m-instruct",
|
|
"full_stack": false,
|
|
"dtype": "bfloat16",
|
|
"stages": {
|
|
"1_config": "needs patch: config.text_config",
|
|
"2_load_causal_lm": "ok",
|
|
"3_container": "needs patch: repo wants model.language_model.layers, found model.layers",
|
|
"4_convention": "MISMATCH: second positional is 'attention_mask', adapter forwards position_embeddings positionally",
|
|
"5_surgery": "ok (no shims)",
|
|
"6_identity": "identical",
|
|
"7_write_read": "ok"
|
|
},
|
|
"patches": [],
|
|
"architecture": "GraniteMoeForCausalLM",
|
|
"model_type": "granitemoe",
|
|
"hidden_size": 1536,
|
|
"num_layers": 32,
|
|
"vocab_size": 49155,
|
|
"text_config_present": false,
|
|
"backbone_class": "GraniteMoeForCausalLM",
|
|
"container_path": "model.layers",
|
|
"repo_layer_path_ok": false,
|
|
"layer_forward_params": [
|
|
"hidden_states",
|
|
"attention_mask",
|
|
"past_key_values",
|
|
"position_embeddings",
|
|
"kwargs"
|
|
],
|
|
"second_positional_is_position_embeddings": false,
|
|
"resolved_layer_indices": [
|
|
7,
|
|
15,
|
|
23,
|
|
31
|
|
],
|
|
"shims_needed": [],
|
|
"identity_max_abs_logit_delta": 0.0,
|
|
"memory_state_changed_by": null
|
|
}
|
|
] |