Add files using upload-large-folder tool
Browse files- README.md +3 -0
- bucket_migration_manifest.json +31 -0
- config.json +118 -0
- configuration_ncp_olmo3.py +32 -0
- conversion_manifest.json +271 -0
- merges.txt +0 -0
- model-00001-of-00005.safetensors +3 -0
- model-00002-of-00005.safetensors +3 -0
- model-00003-of-00005.safetensors +3 -0
- model-00004-of-00005.safetensors +3 -0
- model-00005-of-00005.safetensors +3 -0
- model.safetensors.index.json +849 -0
- modeling_ncp_olmo3.py +1232 -0
- special_tokens_map.json +6 -0
- standalone_manifest.json +25 -0
- tokenizer.json +0 -0
- tokenizer_config.json +189 -0
- training_config.yaml +2805 -0
- vocab.json +0 -0
README.md
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ConceptLM NCP_Olmo3_Stage1_StepLast pure-HF checkpoint
|
| 2 |
+
|
| 3 |
+
This artifact stores one shared set of SafeTensors with dedicated Hugging Face projection keys (`q_proj`, `k_proj`, `v_proj`, `gate_proj`, `up_proj`, and `down_proj`). It loads through `AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True)` and through the ConceptLM vLLM backend; no native Megatron-key weight copy is required. See `conversion_manifest.json` for source and lossless key-conversion evidence.
|
bucket_migration_manifest.json
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"checkpoint_step": 1382310,
|
| 3 |
+
"created_at_utc": "2026-07-27T10:10:45.587395+00:00",
|
| 4 |
+
"destination_model_remote": "plmhhdd2:/plm-hhdd2/model_exports/conceptlm_v22_vq_step1382310_hf",
|
| 5 |
+
"external_runtime_required": false,
|
| 6 |
+
"migrate_to_mount_skill": {
|
| 7 |
+
"fallback": "direct_rclone_copy_with_same_config",
|
| 8 |
+
"mount_attempt_on_login": "failed_before_copy_missing_fusermount3",
|
| 9 |
+
"path": "/mnt/shared-storage-gpfs2/plm-gpfs/ylliu/codex/home/skills/migrate-to-mount/SKILL.md"
|
| 10 |
+
},
|
| 11 |
+
"parity_validated": false,
|
| 12 |
+
"post_export_packaging": "standalone_hf_parity_candidate",
|
| 13 |
+
"remote_validation": {
|
| 14 |
+
"adapter_manifest_checked": false,
|
| 15 |
+
"conversion_manifest_checked": true,
|
| 16 |
+
"expected_object_count": 19,
|
| 17 |
+
"rclone_check_size_only": true,
|
| 18 |
+
"standalone_manifest_checked": true
|
| 19 |
+
},
|
| 20 |
+
"rjob_name": "plm-wyx-dddiag-hf14-0727b",
|
| 21 |
+
"scratch_cleanup_after_remote_validation": true,
|
| 22 |
+
"source_checkpoint_bytes": 92679254212,
|
| 23 |
+
"source_checkpoint_metadata_sha256": "326e9f87dd52409025792b714f76d6de840739137fb7583212b7b8207f69186e",
|
| 24 |
+
"source_checkpoint_objects": 259,
|
| 25 |
+
"source_checkpoint_remote": "plmhhdd2:/plm-hhdd2/dddiag256_full_1f85baf_0611113636/iter_1382310",
|
| 26 |
+
"source_checkpoint_step": 1382310,
|
| 27 |
+
"standalone_hf_backend": true,
|
| 28 |
+
"standalone_loader_updated_at": "2026-07-29T04:24:57+00:00",
|
| 29 |
+
"status": "BUCKET_HF_EXPORT_OK",
|
| 30 |
+
"transport": "rclone_copy"
|
| 31 |
+
}
|
config.json
ADDED
|
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bias_linear": false,
|
| 3 |
+
"add_qkv_bias": false,
|
| 4 |
+
"apply_residual_connection_post_layernorm": false,
|
| 5 |
+
"architectures": [
|
| 6 |
+
"NCPOlmo3ForCausalLM"
|
| 7 |
+
],
|
| 8 |
+
"attention_dropout": 0.0,
|
| 9 |
+
"auto_map": {
|
| 10 |
+
"AutoConfig": "configuration_ncp_olmo3.NCPOlmo3Config",
|
| 11 |
+
"AutoModelForCausalLM": "modeling_ncp_olmo3.NCPOlmo3ForCausalLM"
|
| 12 |
+
},
|
| 13 |
+
"checkpoint_step": 1382310,
|
| 14 |
+
"concept_chunk_size": 4,
|
| 15 |
+
"concept_shift_feature": true,
|
| 16 |
+
"conceptlm_backbone": "olmo3",
|
| 17 |
+
"conceptlm_chunk_dualpath_alpha_init": 0.5,
|
| 18 |
+
"conceptlm_chunk_merge_method": "meanpooling",
|
| 19 |
+
"conceptlm_chunk_size": 4,
|
| 20 |
+
"conceptlm_decoder_layers": 16,
|
| 21 |
+
"conceptlm_enable_chunk_dualpath_smoothing": false,
|
| 22 |
+
"conceptlm_encoder_layers": 16,
|
| 23 |
+
"conceptlm_fusion_alpha_init": 1.0,
|
| 24 |
+
"conceptlm_fusion_norm_alpha_init": 0.1,
|
| 25 |
+
"conceptlm_hlm_attention_mode": "backbone_window",
|
| 26 |
+
"conceptlm_hlm_ffn_hidden_size": null,
|
| 27 |
+
"conceptlm_hlm_loss_weight": 1.0,
|
| 28 |
+
"conceptlm_layer_norm_option": "normed_add",
|
| 29 |
+
"conceptlm_shift_feature": true,
|
| 30 |
+
"conceptlm_special_layers": 8,
|
| 31 |
+
"conceptlm_v21_compile_dd_routes": true,
|
| 32 |
+
"conceptlm_v21_compile_residual_flow_routes": true,
|
| 33 |
+
"conceptlm_v21_concept_read_encoder_first_n": -1,
|
| 34 |
+
"conceptlm_v21_dd_concept_self_dd": true,
|
| 35 |
+
"conceptlm_v21_dd_concept_self_dd_every_n_layers": 1,
|
| 36 |
+
"conceptlm_v21_dd_concept_self_dd_hidden_size": 0,
|
| 37 |
+
"conceptlm_v21_dd_concept_self_dd_use_layernorm": false,
|
| 38 |
+
"conceptlm_v21_dd_encoder_self_dd": true,
|
| 39 |
+
"conceptlm_v21_dd_encoder_self_dd_every_n_layers": 1,
|
| 40 |
+
"conceptlm_v21_dd_encoder_self_dd_hidden_size": 0,
|
| 41 |
+
"conceptlm_v21_dd_encoder_self_dd_use_layernorm": false,
|
| 42 |
+
"conceptlm_v21_dd_self_dd_mode": "dd",
|
| 43 |
+
"conceptlm_v21_dd_two_route_add": true,
|
| 44 |
+
"conceptlm_v21_dd_two_route_add_beta_init": 0.05,
|
| 45 |
+
"conceptlm_v21_dd_two_route_add_concept_hidden_size": 0,
|
| 46 |
+
"conceptlm_v21_dd_two_route_add_concept_route_first_n": -1,
|
| 47 |
+
"conceptlm_v21_dd_two_route_add_concept_source": "final",
|
| 48 |
+
"conceptlm_v21_dd_two_route_add_concept_use_layernorm": true,
|
| 49 |
+
"conceptlm_v21_dd_two_route_add_decoder_hidden_size": 0,
|
| 50 |
+
"conceptlm_v21_dd_two_route_add_decoder_use_layernorm": false,
|
| 51 |
+
"conceptlm_v21_dd_two_route_add_decoder_use_softmax": true,
|
| 52 |
+
"conceptlm_v21_dd_two_route_add_disable_decoder_dd": false,
|
| 53 |
+
"conceptlm_v21_dd_two_route_add_enable_final_concept_route": true,
|
| 54 |
+
"conceptlm_v21_dd_two_route_add_enable_raw_concept_route": false,
|
| 55 |
+
"conceptlm_v21_dd_two_route_add_every_n_layers": 1,
|
| 56 |
+
"conceptlm_v21_dd_two_route_add_use_softmax": false,
|
| 57 |
+
"conceptlm_v21_decoder_read_encoder_first_n": -1,
|
| 58 |
+
"conceptlm_v21_enable_concept_read_encoder": true,
|
| 59 |
+
"conceptlm_v21_enable_decoder_read_concept": true,
|
| 60 |
+
"conceptlm_v21_enable_decoder_read_encoder": true,
|
| 61 |
+
"conceptlm_v21_enable_full_residual_flow": false,
|
| 62 |
+
"conceptlm_v21_final_read_concept_gate": true,
|
| 63 |
+
"conceptlm_v21_final_read_concept_gate_init_final": 0.5,
|
| 64 |
+
"conceptlm_v21_final_read_concept_gate_reg_weight": 0.0,
|
| 65 |
+
"conceptlm_v21_final_read_concept_gate_target_final": 0.5,
|
| 66 |
+
"conceptlm_v21_residual_flow_beta_init": 0.02,
|
| 67 |
+
"conceptlm_v21_residual_flow_route_hidden_size": 0,
|
| 68 |
+
"conceptlm_v21_residual_flow_route_use_softmax": true,
|
| 69 |
+
"conceptlm_v21_residual_flow_shared_source_norm": true,
|
| 70 |
+
"conceptlm_v21_residual_flow_source_use_layernorm": true,
|
| 71 |
+
"conceptlm_v22_vq_codebook_size": 128,
|
| 72 |
+
"conceptlm_v22_vq_commitment_cost": 0.25,
|
| 73 |
+
"conceptlm_v22_vq_detach_hlm_target": true,
|
| 74 |
+
"conceptlm_v22_vq_highlevel_lr_mult": 1.0,
|
| 75 |
+
"conceptlm_v22_vq_hlm_loss_type": "mse",
|
| 76 |
+
"conceptlm_v22_vq_lr_mult": 1.0,
|
| 77 |
+
"conceptlm_v22_vq_merge_mode": "raw_logits",
|
| 78 |
+
"conceptlm_v22_vq_num_codebooks": 32,
|
| 79 |
+
"conceptlm_vq_loss_weight": 1.0,
|
| 80 |
+
"custom_model_loader_required": true,
|
| 81 |
+
"ffn_hidden_size": 11008,
|
| 82 |
+
"fp32_residual_connection": false,
|
| 83 |
+
"gated_linear_unit": true,
|
| 84 |
+
"hidden_dropout": 0.0,
|
| 85 |
+
"hidden_size": 4096,
|
| 86 |
+
"kv_channels": 128,
|
| 87 |
+
"layernorm_epsilon": 1e-06,
|
| 88 |
+
"layernorm_zero_centered_gamma": false,
|
| 89 |
+
"max_position_embeddings": 8192,
|
| 90 |
+
"max_sequence_length": 8192,
|
| 91 |
+
"model_type": "ncp_olmo3",
|
| 92 |
+
"normalization": "RMSNorm",
|
| 93 |
+
"num_attention_heads": 32,
|
| 94 |
+
"num_hidden_layers": 40,
|
| 95 |
+
"num_layers": 32,
|
| 96 |
+
"num_query_groups": 32,
|
| 97 |
+
"params_dtype": "bfloat16",
|
| 98 |
+
"position_embedding_type": "rope",
|
| 99 |
+
"qk_layernorm": true,
|
| 100 |
+
"rotary_base": 500000,
|
| 101 |
+
"rotary_interleaved": false,
|
| 102 |
+
"rotary_percent": 1.0,
|
| 103 |
+
"standalone_hf_backend": true,
|
| 104 |
+
"standalone_hf_backend_status": "parity_candidate",
|
| 105 |
+
"torch_dtype": "bfloat16",
|
| 106 |
+
"transformers_version": "5.8.0",
|
| 107 |
+
"untie_embeddings_and_output_weights": true,
|
| 108 |
+
"use_cache": true,
|
| 109 |
+
"vocab_size": 100278,
|
| 110 |
+
"weight_key_format": "huggingface_state_dict",
|
| 111 |
+
"window_attn_skip_freq": 4,
|
| 112 |
+
"window_size": [
|
| 113 |
+
4096,
|
| 114 |
+
0
|
| 115 |
+
],
|
| 116 |
+
"external_runtime_required": false,
|
| 117 |
+
"runtime_profile": "preserve"
|
| 118 |
+
}
|
configuration_ncp_olmo3.py
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Configuration for the standalone NCP OLMo3 Hugging Face model."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
from transformers import PretrainedConfig
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
class NCPOlmo3Config(PretrainedConfig):
|
| 9 |
+
"""Self-contained inference configuration for NCP OLMo3 checkpoints."""
|
| 10 |
+
|
| 11 |
+
model_type = "ncp_olmo3"
|
| 12 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 13 |
+
|
| 14 |
+
def __init__(self, **kwargs):
|
| 15 |
+
kwargs.setdefault("tie_word_embeddings", False)
|
| 16 |
+
kwargs.setdefault("use_cache", True)
|
| 17 |
+
kwargs.setdefault("is_decoder", True)
|
| 18 |
+
kwargs.setdefault("is_encoder_decoder", False)
|
| 19 |
+
kwargs.setdefault(
|
| 20 |
+
"num_hidden_layers",
|
| 21 |
+
int(kwargs.get("conceptlm_encoder_layers", 0))
|
| 22 |
+
+ int(kwargs.get("conceptlm_special_layers", 0))
|
| 23 |
+
+ int(kwargs.get("conceptlm_decoder_layers", 0)),
|
| 24 |
+
)
|
| 25 |
+
use_cache = bool(kwargs["use_cache"])
|
| 26 |
+
num_hidden_layers = int(kwargs["num_hidden_layers"])
|
| 27 |
+
super().__init__(**kwargs)
|
| 28 |
+
# Transformers 5.x consumes some generation/model-shape kwargs without
|
| 29 |
+
# retaining them on PretrainedConfig. Our forward and generation
|
| 30 |
+
# contracts read these attributes directly, so preserve them explicitly.
|
| 31 |
+
self.use_cache = use_cache
|
| 32 |
+
self.num_hidden_layers = num_hidden_layers
|
conversion_manifest.json
ADDED
|
@@ -0,0 +1,271 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architecture": "NCPOlmo3ForCausalLM",
|
| 3 |
+
"architectures": [
|
| 4 |
+
"NCPOlmo3ForCausalLM"
|
| 5 |
+
],
|
| 6 |
+
"auto_map": {
|
| 7 |
+
"AutoConfig": "configuration_ncp_olmo3.NCPOlmo3Config",
|
| 8 |
+
"AutoModelForCausalLM": "modeling_ncp_olmo3.NCPOlmo3ForCausalLM"
|
| 9 |
+
},
|
| 10 |
+
"checkpoint_step": 1382310,
|
| 11 |
+
"config_sha256": "090ad8898f18d61f21fd506634540fddbfb4cbbb462c8877a2b5b13b08d1e54b",
|
| 12 |
+
"configuration_ncp_olmo3_sha256": "c0282245cba4dd759b3c5feb3d1638e150c14e565641a7dd62bd1d861d497222",
|
| 13 |
+
"created_at_utc": "2026-07-17T13:59:27.348593+00:00",
|
| 14 |
+
"custom_model_loader_required": false,
|
| 15 |
+
"external_runtime_required": false,
|
| 16 |
+
"legacy_architecture": "ConceptLMV22VQForCausalLM",
|
| 17 |
+
"max_shard_bytes": 4294967296,
|
| 18 |
+
"model_type": "ncp_olmo3",
|
| 19 |
+
"modeling_ncp_olmo3_sha256": "cedad85d81dab60c0cb60ce08c0aa940f9074c19ffe14e1175e2e7011aaf3644",
|
| 20 |
+
"parity_validated": false,
|
| 21 |
+
"shards": [
|
| 22 |
+
{
|
| 23 |
+
"file": "model-00001-of-00005.safetensors",
|
| 24 |
+
"tensor_count": 406,
|
| 25 |
+
"file_size": 4267200488,
|
| 26 |
+
"sha256": "db5edee664770d33fd8c857a733ff02edc36b07a32ccb829b6f7ad9b314223af"
|
| 27 |
+
},
|
| 28 |
+
{
|
| 29 |
+
"file": "model-00002-of-00005.safetensors",
|
| 30 |
+
"tensor_count": 119,
|
| 31 |
+
"file_size": 4272273160,
|
| 32 |
+
"sha256": "ad8a6ab4ed71bfcc7175ea1a61cb87db713402f814d1b1a167744b851b642e33"
|
| 33 |
+
},
|
| 34 |
+
{
|
| 35 |
+
"file": "model-00003-of-00005.safetensors",
|
| 36 |
+
"tensor_count": 191,
|
| 37 |
+
"file_size": 4243570472,
|
| 38 |
+
"sha256": "18abcfc27dccc0139ab407d0d91edf334606835de246bba687d587dfa0171470"
|
| 39 |
+
},
|
| 40 |
+
{
|
| 41 |
+
"file": "model-00004-of-00005.safetensors",
|
| 42 |
+
"tensor_count": 125,
|
| 43 |
+
"file_size": 4272306530,
|
| 44 |
+
"sha256": "be473bdccab10029a7d1c5c5ba5d487ee56e839fa1496730225e3ed3fa4fc263"
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"file": "model-00005-of-00005.safetensors",
|
| 48 |
+
"tensor_count": 1,
|
| 49 |
+
"file_size": 821477504,
|
| 50 |
+
"sha256": "4f5ae090db492e2ecaf6f12cf3d8b5b444527aa2c1b2a47370cb3a0206b8ec83"
|
| 51 |
+
}
|
| 52 |
+
],
|
| 53 |
+
"skipped_non_tensor_state_keys": [
|
| 54 |
+
"concept_predictor.hlm_block.layers.0.mlp.linear_fc1._extra_state",
|
| 55 |
+
"concept_predictor.hlm_block.layers.0.mlp.linear_fc2._extra_state",
|
| 56 |
+
"concept_predictor.hlm_block.layers.0.self_attention.linear_proj._extra_state",
|
| 57 |
+
"concept_predictor.hlm_block.layers.0.self_attention.linear_qkv._extra_state",
|
| 58 |
+
"concept_predictor.hlm_block.layers.1.mlp.linear_fc1._extra_state",
|
| 59 |
+
"concept_predictor.hlm_block.layers.1.mlp.linear_fc2._extra_state",
|
| 60 |
+
"concept_predictor.hlm_block.layers.1.self_attention.linear_proj._extra_state",
|
| 61 |
+
"concept_predictor.hlm_block.layers.1.self_attention.linear_qkv._extra_state",
|
| 62 |
+
"concept_predictor.hlm_block.layers.2.mlp.linear_fc1._extra_state",
|
| 63 |
+
"concept_predictor.hlm_block.layers.2.mlp.linear_fc2._extra_state",
|
| 64 |
+
"concept_predictor.hlm_block.layers.2.self_attention.linear_proj._extra_state",
|
| 65 |
+
"concept_predictor.hlm_block.layers.2.self_attention.linear_qkv._extra_state",
|
| 66 |
+
"concept_predictor.hlm_block.layers.3.mlp.linear_fc1._extra_state",
|
| 67 |
+
"concept_predictor.hlm_block.layers.3.mlp.linear_fc2._extra_state",
|
| 68 |
+
"concept_predictor.hlm_block.layers.3.self_attention.linear_proj._extra_state",
|
| 69 |
+
"concept_predictor.hlm_block.layers.3.self_attention.linear_qkv._extra_state",
|
| 70 |
+
"concept_predictor.hlm_block.layers.4.mlp.linear_fc1._extra_state",
|
| 71 |
+
"concept_predictor.hlm_block.layers.4.mlp.linear_fc2._extra_state",
|
| 72 |
+
"concept_predictor.hlm_block.layers.4.self_attention.linear_proj._extra_state",
|
| 73 |
+
"concept_predictor.hlm_block.layers.4.self_attention.linear_qkv._extra_state",
|
| 74 |
+
"concept_predictor.hlm_block.layers.5.mlp.linear_fc1._extra_state",
|
| 75 |
+
"concept_predictor.hlm_block.layers.5.mlp.linear_fc2._extra_state",
|
| 76 |
+
"concept_predictor.hlm_block.layers.5.self_attention.linear_proj._extra_state",
|
| 77 |
+
"concept_predictor.hlm_block.layers.5.self_attention.linear_qkv._extra_state",
|
| 78 |
+
"concept_predictor.hlm_block.layers.6.mlp.linear_fc1._extra_state",
|
| 79 |
+
"concept_predictor.hlm_block.layers.6.mlp.linear_fc2._extra_state",
|
| 80 |
+
"concept_predictor.hlm_block.layers.6.self_attention.linear_proj._extra_state",
|
| 81 |
+
"concept_predictor.hlm_block.layers.6.self_attention.linear_qkv._extra_state",
|
| 82 |
+
"concept_predictor.hlm_block.layers.7.mlp.linear_fc1._extra_state",
|
| 83 |
+
"concept_predictor.hlm_block.layers.7.mlp.linear_fc2._extra_state",
|
| 84 |
+
"concept_predictor.hlm_block.layers.7.self_attention.linear_proj._extra_state",
|
| 85 |
+
"concept_predictor.hlm_block.layers.7.self_attention.linear_qkv._extra_state",
|
| 86 |
+
"decoder.layers.0.mlp.linear_fc1._extra_state",
|
| 87 |
+
"decoder.layers.0.mlp.linear_fc2._extra_state",
|
| 88 |
+
"decoder.layers.0.self_attention.linear_proj._extra_state",
|
| 89 |
+
"decoder.layers.0.self_attention.linear_qkv._extra_state",
|
| 90 |
+
"decoder.layers.1.mlp.linear_fc1._extra_state",
|
| 91 |
+
"decoder.layers.1.mlp.linear_fc2._extra_state",
|
| 92 |
+
"decoder.layers.1.self_attention.linear_proj._extra_state",
|
| 93 |
+
"decoder.layers.1.self_attention.linear_qkv._extra_state",
|
| 94 |
+
"decoder.layers.10.mlp.linear_fc1._extra_state",
|
| 95 |
+
"decoder.layers.10.mlp.linear_fc2._extra_state",
|
| 96 |
+
"decoder.layers.10.self_attention.linear_proj._extra_state",
|
| 97 |
+
"decoder.layers.10.self_attention.linear_qkv._extra_state",
|
| 98 |
+
"decoder.layers.11.mlp.linear_fc1._extra_state",
|
| 99 |
+
"decoder.layers.11.mlp.linear_fc2._extra_state",
|
| 100 |
+
"decoder.layers.11.self_attention.linear_proj._extra_state",
|
| 101 |
+
"decoder.layers.11.self_attention.linear_qkv._extra_state",
|
| 102 |
+
"decoder.layers.12.mlp.linear_fc1._extra_state",
|
| 103 |
+
"decoder.layers.12.mlp.linear_fc2._extra_state",
|
| 104 |
+
"decoder.layers.12.self_attention.linear_proj._extra_state",
|
| 105 |
+
"decoder.layers.12.self_attention.linear_qkv._extra_state",
|
| 106 |
+
"decoder.layers.13.mlp.linear_fc1._extra_state",
|
| 107 |
+
"decoder.layers.13.mlp.linear_fc2._extra_state",
|
| 108 |
+
"decoder.layers.13.self_attention.linear_proj._extra_state",
|
| 109 |
+
"decoder.layers.13.self_attention.linear_qkv._extra_state",
|
| 110 |
+
"decoder.layers.14.mlp.linear_fc1._extra_state",
|
| 111 |
+
"decoder.layers.14.mlp.linear_fc2._extra_state",
|
| 112 |
+
"decoder.layers.14.self_attention.linear_proj._extra_state",
|
| 113 |
+
"decoder.layers.14.self_attention.linear_qkv._extra_state",
|
| 114 |
+
"decoder.layers.15.mlp.linear_fc1._extra_state",
|
| 115 |
+
"decoder.layers.15.mlp.linear_fc2._extra_state",
|
| 116 |
+
"decoder.layers.15.self_attention.linear_proj._extra_state",
|
| 117 |
+
"decoder.layers.15.self_attention.linear_qkv._extra_state",
|
| 118 |
+
"decoder.layers.2.mlp.linear_fc1._extra_state",
|
| 119 |
+
"decoder.layers.2.mlp.linear_fc2._extra_state",
|
| 120 |
+
"decoder.layers.2.self_attention.linear_proj._extra_state",
|
| 121 |
+
"decoder.layers.2.self_attention.linear_qkv._extra_state",
|
| 122 |
+
"decoder.layers.3.mlp.linear_fc1._extra_state",
|
| 123 |
+
"decoder.layers.3.mlp.linear_fc2._extra_state",
|
| 124 |
+
"decoder.layers.3.self_attention.linear_proj._extra_state",
|
| 125 |
+
"decoder.layers.3.self_attention.linear_qkv._extra_state",
|
| 126 |
+
"decoder.layers.4.mlp.linear_fc1._extra_state",
|
| 127 |
+
"decoder.layers.4.mlp.linear_fc2._extra_state",
|
| 128 |
+
"decoder.layers.4.self_attention.linear_proj._extra_state",
|
| 129 |
+
"decoder.layers.4.self_attention.linear_qkv._extra_state",
|
| 130 |
+
"decoder.layers.5.mlp.linear_fc1._extra_state",
|
| 131 |
+
"decoder.layers.5.mlp.linear_fc2._extra_state",
|
| 132 |
+
"decoder.layers.5.self_attention.linear_proj._extra_state",
|
| 133 |
+
"decoder.layers.5.self_attention.linear_qkv._extra_state",
|
| 134 |
+
"decoder.layers.6.mlp.linear_fc1._extra_state",
|
| 135 |
+
"decoder.layers.6.mlp.linear_fc2._extra_state",
|
| 136 |
+
"decoder.layers.6.self_attention.linear_proj._extra_state",
|
| 137 |
+
"decoder.layers.6.self_attention.linear_qkv._extra_state",
|
| 138 |
+
"decoder.layers.7.mlp.linear_fc1._extra_state",
|
| 139 |
+
"decoder.layers.7.mlp.linear_fc2._extra_state",
|
| 140 |
+
"decoder.layers.7.self_attention.linear_proj._extra_state",
|
| 141 |
+
"decoder.layers.7.self_attention.linear_qkv._extra_state",
|
| 142 |
+
"decoder.layers.8.mlp.linear_fc1._extra_state",
|
| 143 |
+
"decoder.layers.8.mlp.linear_fc2._extra_state",
|
| 144 |
+
"decoder.layers.8.self_attention.linear_proj._extra_state",
|
| 145 |
+
"decoder.layers.8.self_attention.linear_qkv._extra_state",
|
| 146 |
+
"decoder.layers.9.mlp.linear_fc1._extra_state",
|
| 147 |
+
"decoder.layers.9.mlp.linear_fc2._extra_state",
|
| 148 |
+
"decoder.layers.9.self_attention.linear_proj._extra_state",
|
| 149 |
+
"decoder.layers.9.self_attention.linear_qkv._extra_state",
|
| 150 |
+
"encoder.layers.0.mlp.linear_fc1._extra_state",
|
| 151 |
+
"encoder.layers.0.mlp.linear_fc2._extra_state",
|
| 152 |
+
"encoder.layers.0.self_attention.linear_proj._extra_state",
|
| 153 |
+
"encoder.layers.0.self_attention.linear_qkv._extra_state",
|
| 154 |
+
"encoder.layers.1.mlp.linear_fc1._extra_state",
|
| 155 |
+
"encoder.layers.1.mlp.linear_fc2._extra_state",
|
| 156 |
+
"encoder.layers.1.self_attention.linear_proj._extra_state",
|
| 157 |
+
"encoder.layers.1.self_attention.linear_qkv._extra_state",
|
| 158 |
+
"encoder.layers.10.mlp.linear_fc1._extra_state",
|
| 159 |
+
"encoder.layers.10.mlp.linear_fc2._extra_state",
|
| 160 |
+
"encoder.layers.10.self_attention.linear_proj._extra_state",
|
| 161 |
+
"encoder.layers.10.self_attention.linear_qkv._extra_state",
|
| 162 |
+
"encoder.layers.11.mlp.linear_fc1._extra_state",
|
| 163 |
+
"encoder.layers.11.mlp.linear_fc2._extra_state",
|
| 164 |
+
"encoder.layers.11.self_attention.linear_proj._extra_state",
|
| 165 |
+
"encoder.layers.11.self_attention.linear_qkv._extra_state",
|
| 166 |
+
"encoder.layers.12.mlp.linear_fc1._extra_state",
|
| 167 |
+
"encoder.layers.12.mlp.linear_fc2._extra_state",
|
| 168 |
+
"encoder.layers.12.self_attention.linear_proj._extra_state",
|
| 169 |
+
"encoder.layers.12.self_attention.linear_qkv._extra_state",
|
| 170 |
+
"encoder.layers.13.mlp.linear_fc1._extra_state",
|
| 171 |
+
"encoder.layers.13.mlp.linear_fc2._extra_state",
|
| 172 |
+
"encoder.layers.13.self_attention.linear_proj._extra_state",
|
| 173 |
+
"encoder.layers.13.self_attention.linear_qkv._extra_state",
|
| 174 |
+
"encoder.layers.14.mlp.linear_fc1._extra_state",
|
| 175 |
+
"encoder.layers.14.mlp.linear_fc2._extra_state",
|
| 176 |
+
"encoder.layers.14.self_attention.linear_proj._extra_state",
|
| 177 |
+
"encoder.layers.14.self_attention.linear_qkv._extra_state",
|
| 178 |
+
"encoder.layers.15.mlp.linear_fc1._extra_state",
|
| 179 |
+
"encoder.layers.15.mlp.linear_fc2._extra_state",
|
| 180 |
+
"encoder.layers.15.self_attention.linear_proj._extra_state",
|
| 181 |
+
"encoder.layers.15.self_attention.linear_qkv._extra_state",
|
| 182 |
+
"encoder.layers.2.mlp.linear_fc1._extra_state",
|
| 183 |
+
"encoder.layers.2.mlp.linear_fc2._extra_state",
|
| 184 |
+
"encoder.layers.2.self_attention.linear_proj._extra_state",
|
| 185 |
+
"encoder.layers.2.self_attention.linear_qkv._extra_state",
|
| 186 |
+
"encoder.layers.3.mlp.linear_fc1._extra_state",
|
| 187 |
+
"encoder.layers.3.mlp.linear_fc2._extra_state",
|
| 188 |
+
"encoder.layers.3.self_attention.linear_proj._extra_state",
|
| 189 |
+
"encoder.layers.3.self_attention.linear_qkv._extra_state",
|
| 190 |
+
"encoder.layers.4.mlp.linear_fc1._extra_state",
|
| 191 |
+
"encoder.layers.4.mlp.linear_fc2._extra_state",
|
| 192 |
+
"encoder.layers.4.self_attention.linear_proj._extra_state",
|
| 193 |
+
"encoder.layers.4.self_attention.linear_qkv._extra_state",
|
| 194 |
+
"encoder.layers.5.mlp.linear_fc1._extra_state",
|
| 195 |
+
"encoder.layers.5.mlp.linear_fc2._extra_state",
|
| 196 |
+
"encoder.layers.5.self_attention.linear_proj._extra_state",
|
| 197 |
+
"encoder.layers.5.self_attention.linear_qkv._extra_state",
|
| 198 |
+
"encoder.layers.6.mlp.linear_fc1._extra_state",
|
| 199 |
+
"encoder.layers.6.mlp.linear_fc2._extra_state",
|
| 200 |
+
"encoder.layers.6.self_attention.linear_proj._extra_state",
|
| 201 |
+
"encoder.layers.6.self_attention.linear_qkv._extra_state",
|
| 202 |
+
"encoder.layers.7.mlp.linear_fc1._extra_state",
|
| 203 |
+
"encoder.layers.7.mlp.linear_fc2._extra_state",
|
| 204 |
+
"encoder.layers.7.self_attention.linear_proj._extra_state",
|
| 205 |
+
"encoder.layers.7.self_attention.linear_qkv._extra_state",
|
| 206 |
+
"encoder.layers.8.mlp.linear_fc1._extra_state",
|
| 207 |
+
"encoder.layers.8.mlp.linear_fc2._extra_state",
|
| 208 |
+
"encoder.layers.8.self_attention.linear_proj._extra_state",
|
| 209 |
+
"encoder.layers.8.self_attention.linear_qkv._extra_state",
|
| 210 |
+
"encoder.layers.9.mlp.linear_fc1._extra_state",
|
| 211 |
+
"encoder.layers.9.mlp.linear_fc2._extra_state",
|
| 212 |
+
"encoder.layers.9.self_attention.linear_proj._extra_state",
|
| 213 |
+
"encoder.layers.9.self_attention.linear_qkv._extra_state",
|
| 214 |
+
"output_layer._extra_state"
|
| 215 |
+
],
|
| 216 |
+
"source_checkpoint_metadata": {
|
| 217 |
+
"mtime_ns": 1784259228787584230,
|
| 218 |
+
"path": "/mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/checkpoints/iter_1382310/.metadata",
|
| 219 |
+
"sha256": "326e9f87dd52409025792b714f76d6de840739137fb7583212b7b8207f69186e",
|
| 220 |
+
"size": 754932
|
| 221 |
+
},
|
| 222 |
+
"source_checkpoint_mutated": false,
|
| 223 |
+
"source_checkpoint_root": "/mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/checkpoints",
|
| 224 |
+
"source_checkpoint_step": 1382310,
|
| 225 |
+
"standalone_hf_backend": true,
|
| 226 |
+
"standalone_hf_backend_status": "parity_candidate",
|
| 227 |
+
"standalone_loader_updated_at": "2026-07-29T04:30:13+00:00",
|
| 228 |
+
"status": "PURE_HF_ARTIFACT_EXPORT_OK",
|
| 229 |
+
"tensor_count": 842,
|
| 230 |
+
"tokenizer_files": [
|
| 231 |
+
"merges.txt",
|
| 232 |
+
"special_tokens_map.json",
|
| 233 |
+
"tokenizer.json",
|
| 234 |
+
"tokenizer_config.json",
|
| 235 |
+
"vocab.json"
|
| 236 |
+
],
|
| 237 |
+
"tokenizer_source": "/mnt/shared-storage-gpfs2/plm-gpfs/public_data_model/model/Olmo-3-1025-7B-stage1-step1413814",
|
| 238 |
+
"total_tensor_bytes": 17876727586,
|
| 239 |
+
"training_config_source": "/mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/run/c22vq-dddiag256-full-0611113636/wandb/wandb/offline-run-20260611_114027-pao6unja/files/config.yaml",
|
| 240 |
+
"validation": {
|
| 241 |
+
"shard_count": 5,
|
| 242 |
+
"status": "SAFETENSORS_INDEX_VALIDATED",
|
| 243 |
+
"tensor_count": 724
|
| 244 |
+
},
|
| 245 |
+
"model_label": "NCP_Olmo3_Stage1_StepLast",
|
| 246 |
+
"weight_key_format": "huggingface_state_dict",
|
| 247 |
+
"source_native_export": "/mnt/shared-storage-gpfs2/plm-gpfs/ylliu/hf_upload_staging/ncp_olmo3_batch_edit_20260728/edit/NCP_Olmo3_Stage1_StepLast.native-megatron-delete-pending-20260826",
|
| 248 |
+
"source_native_export_retained": false,
|
| 249 |
+
"source_native_index_sha256": "e43092695ecc6f45a949ff78a6eb0a711528ee73e930dd248db642e93906ec53",
|
| 250 |
+
"source_tensor_count": 724,
|
| 251 |
+
"skipped_native_metadata": [
|
| 252 |
+
"concept_predictor.hlm_block.final_layernorm._extra_state",
|
| 253 |
+
"decoder.final_layernorm._extra_state"
|
| 254 |
+
],
|
| 255 |
+
"lossless_roundtrip_tensor_count": 722,
|
| 256 |
+
"inferred_model_structure": {
|
| 257 |
+
"conceptlm_encoder_layers": 16,
|
| 258 |
+
"conceptlm_decoder_layers": 16,
|
| 259 |
+
"conceptlm_special_layers": 8,
|
| 260 |
+
"conceptlm_chunk_size": 4,
|
| 261 |
+
"conceptlm_v22_vq_num_codebooks": 32,
|
| 262 |
+
"conceptlm_v22_vq_codebook_size": 128
|
| 263 |
+
},
|
| 264 |
+
"runtime_profile": "preserve",
|
| 265 |
+
"runtime_profile_values": null,
|
| 266 |
+
"pure_hf_remote_code": {
|
| 267 |
+
"configuration_sha256": "e8c123ec91670d5e9acfb8c4a8ba273efae9a4e233243e621273a09d2c21d229",
|
| 268 |
+
"modeling_sha256": "798b3ae74171384f7948eb5f6579330257b2f284f2f9e93729d3eeedbc696c4d"
|
| 269 |
+
},
|
| 270 |
+
"converted_at_utc": "2026-08-26T15:22:56.758039+00:00"
|
| 271 |
+
}
|
merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
model-00001-of-00005.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:db5edee664770d33fd8c857a733ff02edc36b07a32ccb829b6f7ad9b314223af
|
| 3 |
+
size 4267200488
|
model-00002-of-00005.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ad8a6ab4ed71bfcc7175ea1a61cb87db713402f814d1b1a167744b851b642e33
|
| 3 |
+
size 4272273160
|
model-00003-of-00005.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:18abcfc27dccc0139ab407d0d91edf334606835de246bba687d587dfa0171470
|
| 3 |
+
size 4243570472
|
model-00004-of-00005.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:be473bdccab10029a7d1c5c5ba5d487ee56e839fa1496730225e3ed3fa4fc263
|
| 3 |
+
size 4272306530
|
model-00005-of-00005.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4f5ae090db492e2ecaf6f12cf3d8b5b444527aa2c1b2a47370cb3a0206b8ec83
|
| 3 |
+
size 821477504
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,849 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_size": 17876727586
|
| 4 |
+
},
|
| 5 |
+
"weight_map": {
|
| 6 |
+
"concept_predictor.concept_read_encoder_routes.0.residual_diag": "model-00001-of-00005.safetensors",
|
| 7 |
+
"concept_predictor.concept_read_encoder_routes.0.w1.weight": "model-00001-of-00005.safetensors",
|
| 8 |
+
"concept_predictor.concept_read_encoder_routes.0.w2.weight": "model-00001-of-00005.safetensors",
|
| 9 |
+
"concept_predictor.concept_read_encoder_routes.1.residual_diag": "model-00001-of-00005.safetensors",
|
| 10 |
+
"concept_predictor.concept_read_encoder_routes.1.w1.weight": "model-00001-of-00005.safetensors",
|
| 11 |
+
"concept_predictor.concept_read_encoder_routes.1.w2.weight": "model-00001-of-00005.safetensors",
|
| 12 |
+
"concept_predictor.concept_read_encoder_routes.2.residual_diag": "model-00001-of-00005.safetensors",
|
| 13 |
+
"concept_predictor.concept_read_encoder_routes.2.w1.weight": "model-00001-of-00005.safetensors",
|
| 14 |
+
"concept_predictor.concept_read_encoder_routes.2.w2.weight": "model-00001-of-00005.safetensors",
|
| 15 |
+
"concept_predictor.concept_read_encoder_routes.3.residual_diag": "model-00001-of-00005.safetensors",
|
| 16 |
+
"concept_predictor.concept_read_encoder_routes.3.w1.weight": "model-00001-of-00005.safetensors",
|
| 17 |
+
"concept_predictor.concept_read_encoder_routes.3.w2.weight": "model-00001-of-00005.safetensors",
|
| 18 |
+
"concept_predictor.concept_read_encoder_routes.4.residual_diag": "model-00001-of-00005.safetensors",
|
| 19 |
+
"concept_predictor.concept_read_encoder_routes.4.w1.weight": "model-00001-of-00005.safetensors",
|
| 20 |
+
"concept_predictor.concept_read_encoder_routes.4.w2.weight": "model-00001-of-00005.safetensors",
|
| 21 |
+
"concept_predictor.concept_read_encoder_routes.5.residual_diag": "model-00001-of-00005.safetensors",
|
| 22 |
+
"concept_predictor.concept_read_encoder_routes.5.w1.weight": "model-00001-of-00005.safetensors",
|
| 23 |
+
"concept_predictor.concept_read_encoder_routes.5.w2.weight": "model-00001-of-00005.safetensors",
|
| 24 |
+
"concept_predictor.concept_read_encoder_routes.6.residual_diag": "model-00001-of-00005.safetensors",
|
| 25 |
+
"concept_predictor.concept_read_encoder_routes.6.w1.weight": "model-00001-of-00005.safetensors",
|
| 26 |
+
"concept_predictor.concept_read_encoder_routes.6.w2.weight": "model-00001-of-00005.safetensors",
|
| 27 |
+
"concept_predictor.concept_read_encoder_routes.7.residual_diag": "model-00001-of-00005.safetensors",
|
| 28 |
+
"concept_predictor.concept_read_encoder_routes.7.w1.weight": "model-00001-of-00005.safetensors",
|
| 29 |
+
"concept_predictor.concept_read_encoder_routes.7.w2.weight": "model-00001-of-00005.safetensors",
|
| 30 |
+
"concept_predictor.concept_read_encoder_shared_source_norm.bias": "model-00001-of-00005.safetensors",
|
| 31 |
+
"concept_predictor.concept_read_encoder_shared_source_norm.weight": "model-00001-of-00005.safetensors",
|
| 32 |
+
"concept_predictor.concept_self_dd.depth_dds.0.static_a": "model-00001-of-00005.safetensors",
|
| 33 |
+
"concept_predictor.concept_self_dd.depth_dds.0.w1.weight": "model-00001-of-00005.safetensors",
|
| 34 |
+
"concept_predictor.concept_self_dd.depth_dds.0.w2.weight": "model-00001-of-00005.safetensors",
|
| 35 |
+
"concept_predictor.concept_self_dd.depth_dds.1.static_a": "model-00001-of-00005.safetensors",
|
| 36 |
+
"concept_predictor.concept_self_dd.depth_dds.1.w1.weight": "model-00001-of-00005.safetensors",
|
| 37 |
+
"concept_predictor.concept_self_dd.depth_dds.1.w2.weight": "model-00001-of-00005.safetensors",
|
| 38 |
+
"concept_predictor.concept_self_dd.depth_dds.2.static_a": "model-00001-of-00005.safetensors",
|
| 39 |
+
"concept_predictor.concept_self_dd.depth_dds.2.w1.weight": "model-00001-of-00005.safetensors",
|
| 40 |
+
"concept_predictor.concept_self_dd.depth_dds.2.w2.weight": "model-00001-of-00005.safetensors",
|
| 41 |
+
"concept_predictor.concept_self_dd.depth_dds.3.static_a": "model-00001-of-00005.safetensors",
|
| 42 |
+
"concept_predictor.concept_self_dd.depth_dds.3.w1.weight": "model-00001-of-00005.safetensors",
|
| 43 |
+
"concept_predictor.concept_self_dd.depth_dds.3.w2.weight": "model-00001-of-00005.safetensors",
|
| 44 |
+
"concept_predictor.concept_self_dd.depth_dds.4.static_a": "model-00001-of-00005.safetensors",
|
| 45 |
+
"concept_predictor.concept_self_dd.depth_dds.4.w1.weight": "model-00001-of-00005.safetensors",
|
| 46 |
+
"concept_predictor.concept_self_dd.depth_dds.4.w2.weight": "model-00001-of-00005.safetensors",
|
| 47 |
+
"concept_predictor.concept_self_dd.depth_dds.5.static_a": "model-00001-of-00005.safetensors",
|
| 48 |
+
"concept_predictor.concept_self_dd.depth_dds.5.w1.weight": "model-00001-of-00005.safetensors",
|
| 49 |
+
"concept_predictor.concept_self_dd.depth_dds.5.w2.weight": "model-00001-of-00005.safetensors",
|
| 50 |
+
"concept_predictor.concept_self_dd.depth_dds.6.static_a": "model-00001-of-00005.safetensors",
|
| 51 |
+
"concept_predictor.concept_self_dd.depth_dds.6.w1.weight": "model-00001-of-00005.safetensors",
|
| 52 |
+
"concept_predictor.concept_self_dd.depth_dds.6.w2.weight": "model-00001-of-00005.safetensors",
|
| 53 |
+
"concept_predictor.concept_self_dd.depth_dds.7.static_a": "model-00001-of-00005.safetensors",
|
| 54 |
+
"concept_predictor.concept_self_dd.depth_dds.7.w1.weight": "model-00001-of-00005.safetensors",
|
| 55 |
+
"concept_predictor.concept_self_dd.depth_dds.7.w2.weight": "model-00001-of-00005.safetensors",
|
| 56 |
+
"concept_predictor.hlm_block.norm.weight": "model-00001-of-00005.safetensors",
|
| 57 |
+
"concept_predictor.hlm_block.layers.0.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 58 |
+
"concept_predictor.hlm_block.layers.0.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 59 |
+
"concept_predictor.hlm_block.layers.0.mlp.down_proj.weight": "model-00001-of-00005.safetensors",
|
| 60 |
+
"concept_predictor.hlm_block.layers.0.post_attention_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 61 |
+
"concept_predictor.hlm_block.layers.0.post_feedforward_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 62 |
+
"concept_predictor.hlm_block.layers.0.self_attn.k_norm.weight": "model-00001-of-00005.safetensors",
|
| 63 |
+
"concept_predictor.hlm_block.layers.0.self_attn.o_proj.weight": "model-00001-of-00005.safetensors",
|
| 64 |
+
"concept_predictor.hlm_block.layers.0.self_attn.q_proj.weight": "model-00001-of-00005.safetensors",
|
| 65 |
+
"concept_predictor.hlm_block.layers.0.self_attn.k_proj.weight": "model-00001-of-00005.safetensors",
|
| 66 |
+
"concept_predictor.hlm_block.layers.0.self_attn.v_proj.weight": "model-00001-of-00005.safetensors",
|
| 67 |
+
"concept_predictor.hlm_block.layers.0.self_attn.q_norm.weight": "model-00001-of-00005.safetensors",
|
| 68 |
+
"concept_predictor.hlm_block.layers.1.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 69 |
+
"concept_predictor.hlm_block.layers.1.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 70 |
+
"concept_predictor.hlm_block.layers.1.mlp.down_proj.weight": "model-00001-of-00005.safetensors",
|
| 71 |
+
"concept_predictor.hlm_block.layers.1.post_attention_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 72 |
+
"concept_predictor.hlm_block.layers.1.post_feedforward_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 73 |
+
"concept_predictor.hlm_block.layers.1.self_attn.k_norm.weight": "model-00001-of-00005.safetensors",
|
| 74 |
+
"concept_predictor.hlm_block.layers.1.self_attn.o_proj.weight": "model-00001-of-00005.safetensors",
|
| 75 |
+
"concept_predictor.hlm_block.layers.1.self_attn.q_proj.weight": "model-00001-of-00005.safetensors",
|
| 76 |
+
"concept_predictor.hlm_block.layers.1.self_attn.k_proj.weight": "model-00001-of-00005.safetensors",
|
| 77 |
+
"concept_predictor.hlm_block.layers.1.self_attn.v_proj.weight": "model-00001-of-00005.safetensors",
|
| 78 |
+
"concept_predictor.hlm_block.layers.1.self_attn.q_norm.weight": "model-00001-of-00005.safetensors",
|
| 79 |
+
"concept_predictor.hlm_block.layers.2.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 80 |
+
"concept_predictor.hlm_block.layers.2.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 81 |
+
"concept_predictor.hlm_block.layers.2.mlp.down_proj.weight": "model-00001-of-00005.safetensors",
|
| 82 |
+
"concept_predictor.hlm_block.layers.2.post_attention_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 83 |
+
"concept_predictor.hlm_block.layers.2.post_feedforward_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 84 |
+
"concept_predictor.hlm_block.layers.2.self_attn.k_norm.weight": "model-00001-of-00005.safetensors",
|
| 85 |
+
"concept_predictor.hlm_block.layers.2.self_attn.o_proj.weight": "model-00001-of-00005.safetensors",
|
| 86 |
+
"concept_predictor.hlm_block.layers.2.self_attn.q_proj.weight": "model-00001-of-00005.safetensors",
|
| 87 |
+
"concept_predictor.hlm_block.layers.2.self_attn.k_proj.weight": "model-00001-of-00005.safetensors",
|
| 88 |
+
"concept_predictor.hlm_block.layers.2.self_attn.v_proj.weight": "model-00001-of-00005.safetensors",
|
| 89 |
+
"concept_predictor.hlm_block.layers.2.self_attn.q_norm.weight": "model-00001-of-00005.safetensors",
|
| 90 |
+
"concept_predictor.hlm_block.layers.3.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 91 |
+
"concept_predictor.hlm_block.layers.3.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 92 |
+
"concept_predictor.hlm_block.layers.3.mlp.down_proj.weight": "model-00001-of-00005.safetensors",
|
| 93 |
+
"concept_predictor.hlm_block.layers.3.post_attention_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 94 |
+
"concept_predictor.hlm_block.layers.3.post_feedforward_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 95 |
+
"concept_predictor.hlm_block.layers.3.self_attn.k_norm.weight": "model-00001-of-00005.safetensors",
|
| 96 |
+
"concept_predictor.hlm_block.layers.3.self_attn.o_proj.weight": "model-00001-of-00005.safetensors",
|
| 97 |
+
"concept_predictor.hlm_block.layers.3.self_attn.q_proj.weight": "model-00001-of-00005.safetensors",
|
| 98 |
+
"concept_predictor.hlm_block.layers.3.self_attn.k_proj.weight": "model-00001-of-00005.safetensors",
|
| 99 |
+
"concept_predictor.hlm_block.layers.3.self_attn.v_proj.weight": "model-00001-of-00005.safetensors",
|
| 100 |
+
"concept_predictor.hlm_block.layers.3.self_attn.q_norm.weight": "model-00001-of-00005.safetensors",
|
| 101 |
+
"concept_predictor.hlm_block.layers.4.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 102 |
+
"concept_predictor.hlm_block.layers.4.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 103 |
+
"concept_predictor.hlm_block.layers.4.mlp.down_proj.weight": "model-00001-of-00005.safetensors",
|
| 104 |
+
"concept_predictor.hlm_block.layers.4.post_attention_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 105 |
+
"concept_predictor.hlm_block.layers.4.post_feedforward_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 106 |
+
"concept_predictor.hlm_block.layers.4.self_attn.k_norm.weight": "model-00001-of-00005.safetensors",
|
| 107 |
+
"concept_predictor.hlm_block.layers.4.self_attn.o_proj.weight": "model-00001-of-00005.safetensors",
|
| 108 |
+
"concept_predictor.hlm_block.layers.4.self_attn.q_proj.weight": "model-00001-of-00005.safetensors",
|
| 109 |
+
"concept_predictor.hlm_block.layers.4.self_attn.k_proj.weight": "model-00001-of-00005.safetensors",
|
| 110 |
+
"concept_predictor.hlm_block.layers.4.self_attn.v_proj.weight": "model-00001-of-00005.safetensors",
|
| 111 |
+
"concept_predictor.hlm_block.layers.4.self_attn.q_norm.weight": "model-00001-of-00005.safetensors",
|
| 112 |
+
"concept_predictor.hlm_block.layers.5.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 113 |
+
"concept_predictor.hlm_block.layers.5.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 114 |
+
"concept_predictor.hlm_block.layers.5.mlp.down_proj.weight": "model-00001-of-00005.safetensors",
|
| 115 |
+
"concept_predictor.hlm_block.layers.5.post_attention_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 116 |
+
"concept_predictor.hlm_block.layers.5.post_feedforward_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 117 |
+
"concept_predictor.hlm_block.layers.5.self_attn.k_norm.weight": "model-00001-of-00005.safetensors",
|
| 118 |
+
"concept_predictor.hlm_block.layers.5.self_attn.o_proj.weight": "model-00001-of-00005.safetensors",
|
| 119 |
+
"concept_predictor.hlm_block.layers.5.self_attn.q_proj.weight": "model-00001-of-00005.safetensors",
|
| 120 |
+
"concept_predictor.hlm_block.layers.5.self_attn.k_proj.weight": "model-00001-of-00005.safetensors",
|
| 121 |
+
"concept_predictor.hlm_block.layers.5.self_attn.v_proj.weight": "model-00001-of-00005.safetensors",
|
| 122 |
+
"concept_predictor.hlm_block.layers.5.self_attn.q_norm.weight": "model-00001-of-00005.safetensors",
|
| 123 |
+
"concept_predictor.hlm_block.layers.6.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 124 |
+
"concept_predictor.hlm_block.layers.6.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 125 |
+
"concept_predictor.hlm_block.layers.6.mlp.down_proj.weight": "model-00001-of-00005.safetensors",
|
| 126 |
+
"concept_predictor.hlm_block.layers.6.post_attention_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 127 |
+
"concept_predictor.hlm_block.layers.6.post_feedforward_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 128 |
+
"concept_predictor.hlm_block.layers.6.self_attn.k_norm.weight": "model-00001-of-00005.safetensors",
|
| 129 |
+
"concept_predictor.hlm_block.layers.6.self_attn.o_proj.weight": "model-00001-of-00005.safetensors",
|
| 130 |
+
"concept_predictor.hlm_block.layers.6.self_attn.q_proj.weight": "model-00001-of-00005.safetensors",
|
| 131 |
+
"concept_predictor.hlm_block.layers.6.self_attn.k_proj.weight": "model-00001-of-00005.safetensors",
|
| 132 |
+
"concept_predictor.hlm_block.layers.6.self_attn.v_proj.weight": "model-00001-of-00005.safetensors",
|
| 133 |
+
"concept_predictor.hlm_block.layers.6.self_attn.q_norm.weight": "model-00001-of-00005.safetensors",
|
| 134 |
+
"concept_predictor.hlm_block.layers.7.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 135 |
+
"concept_predictor.hlm_block.layers.7.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 136 |
+
"concept_predictor.hlm_block.layers.7.mlp.down_proj.weight": "model-00001-of-00005.safetensors",
|
| 137 |
+
"concept_predictor.hlm_block.layers.7.post_attention_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 138 |
+
"concept_predictor.hlm_block.layers.7.post_feedforward_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 139 |
+
"concept_predictor.hlm_block.layers.7.self_attn.k_norm.weight": "model-00001-of-00005.safetensors",
|
| 140 |
+
"concept_predictor.hlm_block.layers.7.self_attn.o_proj.weight": "model-00001-of-00005.safetensors",
|
| 141 |
+
"concept_predictor.hlm_block.layers.7.self_attn.q_proj.weight": "model-00001-of-00005.safetensors",
|
| 142 |
+
"concept_predictor.hlm_block.layers.7.self_attn.k_proj.weight": "model-00001-of-00005.safetensors",
|
| 143 |
+
"concept_predictor.hlm_block.layers.7.self_attn.v_proj.weight": "model-00001-of-00005.safetensors",
|
| 144 |
+
"concept_predictor.hlm_block.layers.7.self_attn.q_norm.weight": "model-00001-of-00005.safetensors",
|
| 145 |
+
"concept_predictor.prediction_heads.0.bias": "model-00001-of-00005.safetensors",
|
| 146 |
+
"concept_predictor.prediction_heads.0.weight": "model-00001-of-00005.safetensors",
|
| 147 |
+
"concept_predictor.prediction_heads.1.bias": "model-00001-of-00005.safetensors",
|
| 148 |
+
"concept_predictor.prediction_heads.1.weight": "model-00001-of-00005.safetensors",
|
| 149 |
+
"concept_predictor.prediction_heads.10.bias": "model-00001-of-00005.safetensors",
|
| 150 |
+
"concept_predictor.prediction_heads.10.weight": "model-00001-of-00005.safetensors",
|
| 151 |
+
"concept_predictor.prediction_heads.11.bias": "model-00001-of-00005.safetensors",
|
| 152 |
+
"concept_predictor.prediction_heads.11.weight": "model-00001-of-00005.safetensors",
|
| 153 |
+
"concept_predictor.prediction_heads.12.bias": "model-00001-of-00005.safetensors",
|
| 154 |
+
"concept_predictor.prediction_heads.12.weight": "model-00001-of-00005.safetensors",
|
| 155 |
+
"concept_predictor.prediction_heads.13.bias": "model-00001-of-00005.safetensors",
|
| 156 |
+
"concept_predictor.prediction_heads.13.weight": "model-00001-of-00005.safetensors",
|
| 157 |
+
"concept_predictor.prediction_heads.14.bias": "model-00001-of-00005.safetensors",
|
| 158 |
+
"concept_predictor.prediction_heads.14.weight": "model-00001-of-00005.safetensors",
|
| 159 |
+
"concept_predictor.prediction_heads.15.bias": "model-00001-of-00005.safetensors",
|
| 160 |
+
"concept_predictor.prediction_heads.15.weight": "model-00001-of-00005.safetensors",
|
| 161 |
+
"concept_predictor.prediction_heads.16.bias": "model-00001-of-00005.safetensors",
|
| 162 |
+
"concept_predictor.prediction_heads.16.weight": "model-00001-of-00005.safetensors",
|
| 163 |
+
"concept_predictor.prediction_heads.17.bias": "model-00001-of-00005.safetensors",
|
| 164 |
+
"concept_predictor.prediction_heads.17.weight": "model-00001-of-00005.safetensors",
|
| 165 |
+
"concept_predictor.prediction_heads.18.bias": "model-00001-of-00005.safetensors",
|
| 166 |
+
"concept_predictor.prediction_heads.18.weight": "model-00001-of-00005.safetensors",
|
| 167 |
+
"concept_predictor.prediction_heads.19.bias": "model-00001-of-00005.safetensors",
|
| 168 |
+
"concept_predictor.prediction_heads.19.weight": "model-00001-of-00005.safetensors",
|
| 169 |
+
"concept_predictor.prediction_heads.2.bias": "model-00001-of-00005.safetensors",
|
| 170 |
+
"concept_predictor.prediction_heads.2.weight": "model-00001-of-00005.safetensors",
|
| 171 |
+
"concept_predictor.prediction_heads.20.bias": "model-00001-of-00005.safetensors",
|
| 172 |
+
"concept_predictor.prediction_heads.20.weight": "model-00001-of-00005.safetensors",
|
| 173 |
+
"concept_predictor.prediction_heads.21.bias": "model-00001-of-00005.safetensors",
|
| 174 |
+
"concept_predictor.prediction_heads.21.weight": "model-00001-of-00005.safetensors",
|
| 175 |
+
"concept_predictor.prediction_heads.22.bias": "model-00001-of-00005.safetensors",
|
| 176 |
+
"concept_predictor.prediction_heads.22.weight": "model-00001-of-00005.safetensors",
|
| 177 |
+
"concept_predictor.prediction_heads.23.bias": "model-00001-of-00005.safetensors",
|
| 178 |
+
"concept_predictor.prediction_heads.23.weight": "model-00001-of-00005.safetensors",
|
| 179 |
+
"concept_predictor.prediction_heads.24.bias": "model-00001-of-00005.safetensors",
|
| 180 |
+
"concept_predictor.prediction_heads.24.weight": "model-00001-of-00005.safetensors",
|
| 181 |
+
"concept_predictor.prediction_heads.25.bias": "model-00001-of-00005.safetensors",
|
| 182 |
+
"concept_predictor.prediction_heads.25.weight": "model-00001-of-00005.safetensors",
|
| 183 |
+
"concept_predictor.prediction_heads.26.bias": "model-00001-of-00005.safetensors",
|
| 184 |
+
"concept_predictor.prediction_heads.26.weight": "model-00001-of-00005.safetensors",
|
| 185 |
+
"concept_predictor.prediction_heads.27.bias": "model-00001-of-00005.safetensors",
|
| 186 |
+
"concept_predictor.prediction_heads.27.weight": "model-00001-of-00005.safetensors",
|
| 187 |
+
"concept_predictor.prediction_heads.28.bias": "model-00001-of-00005.safetensors",
|
| 188 |
+
"concept_predictor.prediction_heads.28.weight": "model-00001-of-00005.safetensors",
|
| 189 |
+
"concept_predictor.prediction_heads.29.bias": "model-00001-of-00005.safetensors",
|
| 190 |
+
"concept_predictor.prediction_heads.29.weight": "model-00001-of-00005.safetensors",
|
| 191 |
+
"concept_predictor.prediction_heads.3.bias": "model-00001-of-00005.safetensors",
|
| 192 |
+
"concept_predictor.prediction_heads.3.weight": "model-00001-of-00005.safetensors",
|
| 193 |
+
"concept_predictor.prediction_heads.30.bias": "model-00001-of-00005.safetensors",
|
| 194 |
+
"concept_predictor.prediction_heads.30.weight": "model-00001-of-00005.safetensors",
|
| 195 |
+
"concept_predictor.prediction_heads.31.bias": "model-00001-of-00005.safetensors",
|
| 196 |
+
"concept_predictor.prediction_heads.31.weight": "model-00001-of-00005.safetensors",
|
| 197 |
+
"concept_predictor.prediction_heads.4.bias": "model-00001-of-00005.safetensors",
|
| 198 |
+
"concept_predictor.prediction_heads.4.weight": "model-00001-of-00005.safetensors",
|
| 199 |
+
"concept_predictor.prediction_heads.5.bias": "model-00001-of-00005.safetensors",
|
| 200 |
+
"concept_predictor.prediction_heads.5.weight": "model-00001-of-00005.safetensors",
|
| 201 |
+
"concept_predictor.prediction_heads.6.bias": "model-00001-of-00005.safetensors",
|
| 202 |
+
"concept_predictor.prediction_heads.6.weight": "model-00001-of-00005.safetensors",
|
| 203 |
+
"concept_predictor.prediction_heads.7.bias": "model-00001-of-00005.safetensors",
|
| 204 |
+
"concept_predictor.prediction_heads.7.weight": "model-00001-of-00005.safetensors",
|
| 205 |
+
"concept_predictor.prediction_heads.8.bias": "model-00001-of-00005.safetensors",
|
| 206 |
+
"concept_predictor.prediction_heads.8.weight": "model-00001-of-00005.safetensors",
|
| 207 |
+
"concept_predictor.prediction_heads.9.bias": "model-00001-of-00005.safetensors",
|
| 208 |
+
"concept_predictor.prediction_heads.9.weight": "model-00001-of-00005.safetensors",
|
| 209 |
+
"concept_quantizer.codebook.0": "model-00001-of-00005.safetensors",
|
| 210 |
+
"concept_quantizer.codebook.1": "model-00001-of-00005.safetensors",
|
| 211 |
+
"concept_quantizer.codebook.10": "model-00001-of-00005.safetensors",
|
| 212 |
+
"concept_quantizer.codebook.11": "model-00001-of-00005.safetensors",
|
| 213 |
+
"concept_quantizer.codebook.12": "model-00001-of-00005.safetensors",
|
| 214 |
+
"concept_quantizer.codebook.13": "model-00001-of-00005.safetensors",
|
| 215 |
+
"concept_quantizer.codebook.14": "model-00001-of-00005.safetensors",
|
| 216 |
+
"concept_quantizer.codebook.15": "model-00001-of-00005.safetensors",
|
| 217 |
+
"concept_quantizer.codebook.16": "model-00001-of-00005.safetensors",
|
| 218 |
+
"concept_quantizer.codebook.17": "model-00001-of-00005.safetensors",
|
| 219 |
+
"concept_quantizer.codebook.18": "model-00001-of-00005.safetensors",
|
| 220 |
+
"concept_quantizer.codebook.19": "model-00001-of-00005.safetensors",
|
| 221 |
+
"concept_quantizer.codebook.2": "model-00001-of-00005.safetensors",
|
| 222 |
+
"concept_quantizer.codebook.20": "model-00001-of-00005.safetensors",
|
| 223 |
+
"concept_quantizer.codebook.21": "model-00001-of-00005.safetensors",
|
| 224 |
+
"concept_quantizer.codebook.22": "model-00001-of-00005.safetensors",
|
| 225 |
+
"concept_quantizer.codebook.23": "model-00001-of-00005.safetensors",
|
| 226 |
+
"concept_quantizer.codebook.24": "model-00001-of-00005.safetensors",
|
| 227 |
+
"concept_quantizer.codebook.25": "model-00001-of-00005.safetensors",
|
| 228 |
+
"concept_quantizer.codebook.26": "model-00001-of-00005.safetensors",
|
| 229 |
+
"concept_quantizer.codebook.27": "model-00001-of-00005.safetensors",
|
| 230 |
+
"concept_quantizer.codebook.28": "model-00001-of-00005.safetensors",
|
| 231 |
+
"concept_quantizer.codebook.29": "model-00001-of-00005.safetensors",
|
| 232 |
+
"concept_quantizer.codebook.3": "model-00001-of-00005.safetensors",
|
| 233 |
+
"concept_quantizer.codebook.30": "model-00001-of-00005.safetensors",
|
| 234 |
+
"concept_quantizer.codebook.31": "model-00001-of-00005.safetensors",
|
| 235 |
+
"concept_quantizer.codebook.4": "model-00001-of-00005.safetensors",
|
| 236 |
+
"concept_quantizer.codebook.5": "model-00001-of-00005.safetensors",
|
| 237 |
+
"concept_quantizer.codebook.6": "model-00001-of-00005.safetensors",
|
| 238 |
+
"concept_quantizer.codebook.7": "model-00001-of-00005.safetensors",
|
| 239 |
+
"concept_quantizer.codebook.8": "model-00001-of-00005.safetensors",
|
| 240 |
+
"concept_quantizer.codebook.9": "model-00001-of-00005.safetensors",
|
| 241 |
+
"concept_vq_input_norm.bias": "model-00001-of-00005.safetensors",
|
| 242 |
+
"concept_vq_input_norm.weight": "model-00001-of-00005.safetensors",
|
| 243 |
+
"dd_encoder_self_dd.depth_dds.0.static_a": "model-00001-of-00005.safetensors",
|
| 244 |
+
"dd_encoder_self_dd.depth_dds.0.w1.weight": "model-00001-of-00005.safetensors",
|
| 245 |
+
"dd_encoder_self_dd.depth_dds.0.w2.weight": "model-00001-of-00005.safetensors",
|
| 246 |
+
"dd_encoder_self_dd.depth_dds.1.static_a": "model-00001-of-00005.safetensors",
|
| 247 |
+
"dd_encoder_self_dd.depth_dds.1.w1.weight": "model-00001-of-00005.safetensors",
|
| 248 |
+
"dd_encoder_self_dd.depth_dds.1.w2.weight": "model-00001-of-00005.safetensors",
|
| 249 |
+
"dd_encoder_self_dd.depth_dds.10.static_a": "model-00001-of-00005.safetensors",
|
| 250 |
+
"dd_encoder_self_dd.depth_dds.10.w1.weight": "model-00001-of-00005.safetensors",
|
| 251 |
+
"dd_encoder_self_dd.depth_dds.10.w2.weight": "model-00001-of-00005.safetensors",
|
| 252 |
+
"dd_encoder_self_dd.depth_dds.11.static_a": "model-00001-of-00005.safetensors",
|
| 253 |
+
"dd_encoder_self_dd.depth_dds.11.w1.weight": "model-00001-of-00005.safetensors",
|
| 254 |
+
"dd_encoder_self_dd.depth_dds.11.w2.weight": "model-00001-of-00005.safetensors",
|
| 255 |
+
"dd_encoder_self_dd.depth_dds.12.static_a": "model-00001-of-00005.safetensors",
|
| 256 |
+
"dd_encoder_self_dd.depth_dds.12.w1.weight": "model-00001-of-00005.safetensors",
|
| 257 |
+
"dd_encoder_self_dd.depth_dds.12.w2.weight": "model-00001-of-00005.safetensors",
|
| 258 |
+
"dd_encoder_self_dd.depth_dds.13.static_a": "model-00001-of-00005.safetensors",
|
| 259 |
+
"dd_encoder_self_dd.depth_dds.13.w1.weight": "model-00001-of-00005.safetensors",
|
| 260 |
+
"dd_encoder_self_dd.depth_dds.13.w2.weight": "model-00001-of-00005.safetensors",
|
| 261 |
+
"dd_encoder_self_dd.depth_dds.14.static_a": "model-00001-of-00005.safetensors",
|
| 262 |
+
"dd_encoder_self_dd.depth_dds.14.w1.weight": "model-00001-of-00005.safetensors",
|
| 263 |
+
"dd_encoder_self_dd.depth_dds.14.w2.weight": "model-00001-of-00005.safetensors",
|
| 264 |
+
"dd_encoder_self_dd.depth_dds.15.static_a": "model-00001-of-00005.safetensors",
|
| 265 |
+
"dd_encoder_self_dd.depth_dds.15.w1.weight": "model-00001-of-00005.safetensors",
|
| 266 |
+
"dd_encoder_self_dd.depth_dds.15.w2.weight": "model-00001-of-00005.safetensors",
|
| 267 |
+
"dd_encoder_self_dd.depth_dds.2.static_a": "model-00001-of-00005.safetensors",
|
| 268 |
+
"dd_encoder_self_dd.depth_dds.2.w1.weight": "model-00001-of-00005.safetensors",
|
| 269 |
+
"dd_encoder_self_dd.depth_dds.2.w2.weight": "model-00001-of-00005.safetensors",
|
| 270 |
+
"dd_encoder_self_dd.depth_dds.3.static_a": "model-00001-of-00005.safetensors",
|
| 271 |
+
"dd_encoder_self_dd.depth_dds.3.w1.weight": "model-00001-of-00005.safetensors",
|
| 272 |
+
"dd_encoder_self_dd.depth_dds.3.w2.weight": "model-00001-of-00005.safetensors",
|
| 273 |
+
"dd_encoder_self_dd.depth_dds.4.static_a": "model-00001-of-00005.safetensors",
|
| 274 |
+
"dd_encoder_self_dd.depth_dds.4.w1.weight": "model-00001-of-00005.safetensors",
|
| 275 |
+
"dd_encoder_self_dd.depth_dds.4.w2.weight": "model-00001-of-00005.safetensors",
|
| 276 |
+
"dd_encoder_self_dd.depth_dds.5.static_a": "model-00001-of-00005.safetensors",
|
| 277 |
+
"dd_encoder_self_dd.depth_dds.5.w1.weight": "model-00001-of-00005.safetensors",
|
| 278 |
+
"dd_encoder_self_dd.depth_dds.5.w2.weight": "model-00001-of-00005.safetensors",
|
| 279 |
+
"dd_encoder_self_dd.depth_dds.6.static_a": "model-00001-of-00005.safetensors",
|
| 280 |
+
"dd_encoder_self_dd.depth_dds.6.w1.weight": "model-00001-of-00005.safetensors",
|
| 281 |
+
"dd_encoder_self_dd.depth_dds.6.w2.weight": "model-00001-of-00005.safetensors",
|
| 282 |
+
"dd_encoder_self_dd.depth_dds.7.static_a": "model-00001-of-00005.safetensors",
|
| 283 |
+
"dd_encoder_self_dd.depth_dds.7.w1.weight": "model-00001-of-00005.safetensors",
|
| 284 |
+
"dd_encoder_self_dd.depth_dds.7.w2.weight": "model-00001-of-00005.safetensors",
|
| 285 |
+
"dd_encoder_self_dd.depth_dds.8.static_a": "model-00001-of-00005.safetensors",
|
| 286 |
+
"dd_encoder_self_dd.depth_dds.8.w1.weight": "model-00001-of-00005.safetensors",
|
| 287 |
+
"dd_encoder_self_dd.depth_dds.8.w2.weight": "model-00001-of-00005.safetensors",
|
| 288 |
+
"dd_encoder_self_dd.depth_dds.9.static_a": "model-00001-of-00005.safetensors",
|
| 289 |
+
"dd_encoder_self_dd.depth_dds.9.w1.weight": "model-00001-of-00005.safetensors",
|
| 290 |
+
"dd_encoder_self_dd.depth_dds.9.w2.weight": "model-00001-of-00005.safetensors",
|
| 291 |
+
"dd_two_route_add.concept_routes.0.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 292 |
+
"dd_two_route_add.concept_routes.0.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 293 |
+
"dd_two_route_add.concept_routes.0.final_diag": "model-00001-of-00005.safetensors",
|
| 294 |
+
"dd_two_route_add.concept_routes.1.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 295 |
+
"dd_two_route_add.concept_routes.1.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 296 |
+
"dd_two_route_add.concept_routes.1.final_diag": "model-00001-of-00005.safetensors",
|
| 297 |
+
"dd_two_route_add.concept_routes.10.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 298 |
+
"dd_two_route_add.concept_routes.10.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 299 |
+
"dd_two_route_add.concept_routes.10.final_diag": "model-00001-of-00005.safetensors",
|
| 300 |
+
"dd_two_route_add.concept_routes.11.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 301 |
+
"dd_two_route_add.concept_routes.11.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 302 |
+
"dd_two_route_add.concept_routes.11.final_diag": "model-00001-of-00005.safetensors",
|
| 303 |
+
"dd_two_route_add.concept_routes.12.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 304 |
+
"dd_two_route_add.concept_routes.12.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 305 |
+
"dd_two_route_add.concept_routes.12.final_diag": "model-00001-of-00005.safetensors",
|
| 306 |
+
"dd_two_route_add.concept_routes.13.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 307 |
+
"dd_two_route_add.concept_routes.13.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 308 |
+
"dd_two_route_add.concept_routes.13.final_diag": "model-00001-of-00005.safetensors",
|
| 309 |
+
"dd_two_route_add.concept_routes.14.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 310 |
+
"dd_two_route_add.concept_routes.14.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 311 |
+
"dd_two_route_add.concept_routes.14.final_diag": "model-00001-of-00005.safetensors",
|
| 312 |
+
"dd_two_route_add.concept_routes.15.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 313 |
+
"dd_two_route_add.concept_routes.15.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 314 |
+
"dd_two_route_add.concept_routes.15.final_diag": "model-00001-of-00005.safetensors",
|
| 315 |
+
"dd_two_route_add.concept_routes.2.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 316 |
+
"dd_two_route_add.concept_routes.2.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 317 |
+
"dd_two_route_add.concept_routes.2.final_diag": "model-00001-of-00005.safetensors",
|
| 318 |
+
"dd_two_route_add.concept_routes.3.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 319 |
+
"dd_two_route_add.concept_routes.3.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 320 |
+
"dd_two_route_add.concept_routes.3.final_diag": "model-00001-of-00005.safetensors",
|
| 321 |
+
"dd_two_route_add.concept_routes.4.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 322 |
+
"dd_two_route_add.concept_routes.4.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 323 |
+
"dd_two_route_add.concept_routes.4.final_diag": "model-00001-of-00005.safetensors",
|
| 324 |
+
"dd_two_route_add.concept_routes.5.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 325 |
+
"dd_two_route_add.concept_routes.5.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 326 |
+
"dd_two_route_add.concept_routes.5.final_diag": "model-00001-of-00005.safetensors",
|
| 327 |
+
"dd_two_route_add.concept_routes.6.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 328 |
+
"dd_two_route_add.concept_routes.6.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 329 |
+
"dd_two_route_add.concept_routes.6.final_diag": "model-00001-of-00005.safetensors",
|
| 330 |
+
"dd_two_route_add.concept_routes.7.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 331 |
+
"dd_two_route_add.concept_routes.7.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 332 |
+
"dd_two_route_add.concept_routes.7.final_diag": "model-00001-of-00005.safetensors",
|
| 333 |
+
"dd_two_route_add.concept_routes.8.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 334 |
+
"dd_two_route_add.concept_routes.8.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 335 |
+
"dd_two_route_add.concept_routes.8.final_diag": "model-00001-of-00005.safetensors",
|
| 336 |
+
"dd_two_route_add.concept_routes.9.concept_norm.bias": "model-00001-of-00005.safetensors",
|
| 337 |
+
"dd_two_route_add.concept_routes.9.concept_norm.weight": "model-00001-of-00005.safetensors",
|
| 338 |
+
"dd_two_route_add.concept_routes.9.final_diag": "model-00001-of-00005.safetensors",
|
| 339 |
+
"dd_two_route_add.decoder_dds.0.static_a": "model-00001-of-00005.safetensors",
|
| 340 |
+
"dd_two_route_add.decoder_dds.0.w1.weight": "model-00001-of-00005.safetensors",
|
| 341 |
+
"dd_two_route_add.decoder_dds.0.w2.weight": "model-00001-of-00005.safetensors",
|
| 342 |
+
"dd_two_route_add.decoder_dds.1.static_a": "model-00001-of-00005.safetensors",
|
| 343 |
+
"dd_two_route_add.decoder_dds.1.w1.weight": "model-00001-of-00005.safetensors",
|
| 344 |
+
"dd_two_route_add.decoder_dds.1.w2.weight": "model-00001-of-00005.safetensors",
|
| 345 |
+
"dd_two_route_add.decoder_dds.10.static_a": "model-00001-of-00005.safetensors",
|
| 346 |
+
"dd_two_route_add.decoder_dds.10.w1.weight": "model-00001-of-00005.safetensors",
|
| 347 |
+
"dd_two_route_add.decoder_dds.10.w2.weight": "model-00001-of-00005.safetensors",
|
| 348 |
+
"dd_two_route_add.decoder_dds.11.static_a": "model-00001-of-00005.safetensors",
|
| 349 |
+
"dd_two_route_add.decoder_dds.11.w1.weight": "model-00001-of-00005.safetensors",
|
| 350 |
+
"dd_two_route_add.decoder_dds.11.w2.weight": "model-00001-of-00005.safetensors",
|
| 351 |
+
"dd_two_route_add.decoder_dds.12.static_a": "model-00001-of-00005.safetensors",
|
| 352 |
+
"dd_two_route_add.decoder_dds.12.w1.weight": "model-00001-of-00005.safetensors",
|
| 353 |
+
"dd_two_route_add.decoder_dds.12.w2.weight": "model-00001-of-00005.safetensors",
|
| 354 |
+
"dd_two_route_add.decoder_dds.13.static_a": "model-00001-of-00005.safetensors",
|
| 355 |
+
"dd_two_route_add.decoder_dds.13.w1.weight": "model-00001-of-00005.safetensors",
|
| 356 |
+
"dd_two_route_add.decoder_dds.13.w2.weight": "model-00001-of-00005.safetensors",
|
| 357 |
+
"dd_two_route_add.decoder_dds.14.static_a": "model-00001-of-00005.safetensors",
|
| 358 |
+
"dd_two_route_add.decoder_dds.14.w1.weight": "model-00001-of-00005.safetensors",
|
| 359 |
+
"dd_two_route_add.decoder_dds.14.w2.weight": "model-00001-of-00005.safetensors",
|
| 360 |
+
"dd_two_route_add.decoder_dds.15.static_a": "model-00001-of-00005.safetensors",
|
| 361 |
+
"dd_two_route_add.decoder_dds.15.w1.weight": "model-00001-of-00005.safetensors",
|
| 362 |
+
"dd_two_route_add.decoder_dds.15.w2.weight": "model-00001-of-00005.safetensors",
|
| 363 |
+
"dd_two_route_add.decoder_dds.2.static_a": "model-00001-of-00005.safetensors",
|
| 364 |
+
"dd_two_route_add.decoder_dds.2.w1.weight": "model-00001-of-00005.safetensors",
|
| 365 |
+
"dd_two_route_add.decoder_dds.2.w2.weight": "model-00001-of-00005.safetensors",
|
| 366 |
+
"dd_two_route_add.decoder_dds.3.static_a": "model-00001-of-00005.safetensors",
|
| 367 |
+
"dd_two_route_add.decoder_dds.3.w1.weight": "model-00001-of-00005.safetensors",
|
| 368 |
+
"dd_two_route_add.decoder_dds.3.w2.weight": "model-00001-of-00005.safetensors",
|
| 369 |
+
"dd_two_route_add.decoder_dds.4.static_a": "model-00001-of-00005.safetensors",
|
| 370 |
+
"dd_two_route_add.decoder_dds.4.w1.weight": "model-00001-of-00005.safetensors",
|
| 371 |
+
"dd_two_route_add.decoder_dds.4.w2.weight": "model-00001-of-00005.safetensors",
|
| 372 |
+
"dd_two_route_add.decoder_dds.5.static_a": "model-00001-of-00005.safetensors",
|
| 373 |
+
"dd_two_route_add.decoder_dds.5.w1.weight": "model-00001-of-00005.safetensors",
|
| 374 |
+
"dd_two_route_add.decoder_dds.5.w2.weight": "model-00001-of-00005.safetensors",
|
| 375 |
+
"dd_two_route_add.decoder_dds.6.static_a": "model-00001-of-00005.safetensors",
|
| 376 |
+
"dd_two_route_add.decoder_dds.6.w1.weight": "model-00001-of-00005.safetensors",
|
| 377 |
+
"dd_two_route_add.decoder_dds.6.w2.weight": "model-00001-of-00005.safetensors",
|
| 378 |
+
"dd_two_route_add.decoder_dds.7.static_a": "model-00001-of-00005.safetensors",
|
| 379 |
+
"dd_two_route_add.decoder_dds.7.w1.weight": "model-00001-of-00005.safetensors",
|
| 380 |
+
"dd_two_route_add.decoder_dds.7.w2.weight": "model-00001-of-00005.safetensors",
|
| 381 |
+
"dd_two_route_add.decoder_dds.8.static_a": "model-00001-of-00005.safetensors",
|
| 382 |
+
"dd_two_route_add.decoder_dds.8.w1.weight": "model-00001-of-00005.safetensors",
|
| 383 |
+
"dd_two_route_add.decoder_dds.8.w2.weight": "model-00001-of-00005.safetensors",
|
| 384 |
+
"dd_two_route_add.decoder_dds.9.static_a": "model-00001-of-00005.safetensors",
|
| 385 |
+
"dd_two_route_add.decoder_dds.9.w1.weight": "model-00001-of-00005.safetensors",
|
| 386 |
+
"dd_two_route_add.decoder_dds.9.w2.weight": "model-00001-of-00005.safetensors",
|
| 387 |
+
"decoder.norm.weight": "model-00001-of-00005.safetensors",
|
| 388 |
+
"decoder.layers.0.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 389 |
+
"decoder.layers.0.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 390 |
+
"decoder.layers.0.mlp.down_proj.weight": "model-00001-of-00005.safetensors",
|
| 391 |
+
"decoder.layers.0.post_attention_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 392 |
+
"decoder.layers.0.post_feedforward_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 393 |
+
"decoder.layers.0.self_attn.k_norm.weight": "model-00001-of-00005.safetensors",
|
| 394 |
+
"decoder.layers.0.self_attn.o_proj.weight": "model-00001-of-00005.safetensors",
|
| 395 |
+
"decoder.layers.0.self_attn.q_proj.weight": "model-00001-of-00005.safetensors",
|
| 396 |
+
"decoder.layers.0.self_attn.k_proj.weight": "model-00001-of-00005.safetensors",
|
| 397 |
+
"decoder.layers.0.self_attn.v_proj.weight": "model-00001-of-00005.safetensors",
|
| 398 |
+
"decoder.layers.0.self_attn.q_norm.weight": "model-00001-of-00005.safetensors",
|
| 399 |
+
"decoder.layers.1.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 400 |
+
"decoder.layers.1.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 401 |
+
"decoder.layers.1.mlp.down_proj.weight": "model-00001-of-00005.safetensors",
|
| 402 |
+
"decoder.layers.1.post_attention_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 403 |
+
"decoder.layers.1.post_feedforward_layernorm.weight": "model-00001-of-00005.safetensors",
|
| 404 |
+
"decoder.layers.1.self_attn.k_norm.weight": "model-00001-of-00005.safetensors",
|
| 405 |
+
"decoder.layers.1.self_attn.o_proj.weight": "model-00001-of-00005.safetensors",
|
| 406 |
+
"decoder.layers.1.self_attn.q_proj.weight": "model-00001-of-00005.safetensors",
|
| 407 |
+
"decoder.layers.1.self_attn.k_proj.weight": "model-00001-of-00005.safetensors",
|
| 408 |
+
"decoder.layers.1.self_attn.v_proj.weight": "model-00001-of-00005.safetensors",
|
| 409 |
+
"decoder.layers.1.self_attn.q_norm.weight": "model-00001-of-00005.safetensors",
|
| 410 |
+
"decoder.layers.10.mlp.gate_proj.weight": "model-00001-of-00005.safetensors",
|
| 411 |
+
"decoder.layers.10.mlp.up_proj.weight": "model-00001-of-00005.safetensors",
|
| 412 |
+
"decoder.layers.10.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 413 |
+
"decoder.layers.10.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 414 |
+
"decoder.layers.10.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 415 |
+
"decoder.layers.10.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 416 |
+
"decoder.layers.10.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 417 |
+
"decoder.layers.10.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 418 |
+
"decoder.layers.10.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 419 |
+
"decoder.layers.10.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 420 |
+
"decoder.layers.10.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 421 |
+
"decoder.layers.11.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
|
| 422 |
+
"decoder.layers.11.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
|
| 423 |
+
"decoder.layers.11.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 424 |
+
"decoder.layers.11.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 425 |
+
"decoder.layers.11.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 426 |
+
"decoder.layers.11.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 427 |
+
"decoder.layers.11.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 428 |
+
"decoder.layers.11.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 429 |
+
"decoder.layers.11.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 430 |
+
"decoder.layers.11.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 431 |
+
"decoder.layers.11.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 432 |
+
"decoder.layers.12.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
|
| 433 |
+
"decoder.layers.12.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
|
| 434 |
+
"decoder.layers.12.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 435 |
+
"decoder.layers.12.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 436 |
+
"decoder.layers.12.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 437 |
+
"decoder.layers.12.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 438 |
+
"decoder.layers.12.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 439 |
+
"decoder.layers.12.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 440 |
+
"decoder.layers.12.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 441 |
+
"decoder.layers.12.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 442 |
+
"decoder.layers.12.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 443 |
+
"decoder.layers.13.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
|
| 444 |
+
"decoder.layers.13.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
|
| 445 |
+
"decoder.layers.13.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 446 |
+
"decoder.layers.13.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 447 |
+
"decoder.layers.13.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 448 |
+
"decoder.layers.13.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 449 |
+
"decoder.layers.13.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 450 |
+
"decoder.layers.13.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 451 |
+
"decoder.layers.13.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 452 |
+
"decoder.layers.13.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 453 |
+
"decoder.layers.13.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 454 |
+
"decoder.layers.14.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
|
| 455 |
+
"decoder.layers.14.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
|
| 456 |
+
"decoder.layers.14.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 457 |
+
"decoder.layers.14.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 458 |
+
"decoder.layers.14.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 459 |
+
"decoder.layers.14.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 460 |
+
"decoder.layers.14.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 461 |
+
"decoder.layers.14.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 462 |
+
"decoder.layers.14.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 463 |
+
"decoder.layers.14.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 464 |
+
"decoder.layers.14.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 465 |
+
"decoder.layers.15.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
|
| 466 |
+
"decoder.layers.15.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
|
| 467 |
+
"decoder.layers.15.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 468 |
+
"decoder.layers.15.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 469 |
+
"decoder.layers.15.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 470 |
+
"decoder.layers.15.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 471 |
+
"decoder.layers.15.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 472 |
+
"decoder.layers.15.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 473 |
+
"decoder.layers.15.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 474 |
+
"decoder.layers.15.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 475 |
+
"decoder.layers.15.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 476 |
+
"decoder.layers.2.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
|
| 477 |
+
"decoder.layers.2.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
|
| 478 |
+
"decoder.layers.2.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 479 |
+
"decoder.layers.2.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 480 |
+
"decoder.layers.2.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 481 |
+
"decoder.layers.2.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 482 |
+
"decoder.layers.2.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 483 |
+
"decoder.layers.2.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 484 |
+
"decoder.layers.2.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 485 |
+
"decoder.layers.2.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 486 |
+
"decoder.layers.2.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 487 |
+
"decoder.layers.3.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
|
| 488 |
+
"decoder.layers.3.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
|
| 489 |
+
"decoder.layers.3.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 490 |
+
"decoder.layers.3.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 491 |
+
"decoder.layers.3.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 492 |
+
"decoder.layers.3.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 493 |
+
"decoder.layers.3.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 494 |
+
"decoder.layers.3.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 495 |
+
"decoder.layers.3.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 496 |
+
"decoder.layers.3.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 497 |
+
"decoder.layers.3.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 498 |
+
"decoder.layers.4.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
|
| 499 |
+
"decoder.layers.4.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
|
| 500 |
+
"decoder.layers.4.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 501 |
+
"decoder.layers.4.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 502 |
+
"decoder.layers.4.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 503 |
+
"decoder.layers.4.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 504 |
+
"decoder.layers.4.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 505 |
+
"decoder.layers.4.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 506 |
+
"decoder.layers.4.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 507 |
+
"decoder.layers.4.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 508 |
+
"decoder.layers.4.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 509 |
+
"decoder.layers.5.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
|
| 510 |
+
"decoder.layers.5.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
|
| 511 |
+
"decoder.layers.5.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 512 |
+
"decoder.layers.5.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 513 |
+
"decoder.layers.5.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 514 |
+
"decoder.layers.5.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 515 |
+
"decoder.layers.5.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 516 |
+
"decoder.layers.5.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 517 |
+
"decoder.layers.5.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 518 |
+
"decoder.layers.5.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 519 |
+
"decoder.layers.5.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 520 |
+
"decoder.layers.6.mlp.gate_proj.weight": "model-00002-of-00005.safetensors",
|
| 521 |
+
"decoder.layers.6.mlp.up_proj.weight": "model-00002-of-00005.safetensors",
|
| 522 |
+
"decoder.layers.6.mlp.down_proj.weight": "model-00002-of-00005.safetensors",
|
| 523 |
+
"decoder.layers.6.post_attention_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 524 |
+
"decoder.layers.6.post_feedforward_layernorm.weight": "model-00002-of-00005.safetensors",
|
| 525 |
+
"decoder.layers.6.self_attn.k_norm.weight": "model-00002-of-00005.safetensors",
|
| 526 |
+
"decoder.layers.6.self_attn.o_proj.weight": "model-00002-of-00005.safetensors",
|
| 527 |
+
"decoder.layers.6.self_attn.q_proj.weight": "model-00002-of-00005.safetensors",
|
| 528 |
+
"decoder.layers.6.self_attn.k_proj.weight": "model-00002-of-00005.safetensors",
|
| 529 |
+
"decoder.layers.6.self_attn.v_proj.weight": "model-00002-of-00005.safetensors",
|
| 530 |
+
"decoder.layers.6.self_attn.q_norm.weight": "model-00002-of-00005.safetensors",
|
| 531 |
+
"decoder.layers.7.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
|
| 532 |
+
"decoder.layers.7.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
|
| 533 |
+
"decoder.layers.7.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
|
| 534 |
+
"decoder.layers.7.post_attention_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 535 |
+
"decoder.layers.7.post_feedforward_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 536 |
+
"decoder.layers.7.self_attn.k_norm.weight": "model-00003-of-00005.safetensors",
|
| 537 |
+
"decoder.layers.7.self_attn.o_proj.weight": "model-00003-of-00005.safetensors",
|
| 538 |
+
"decoder.layers.7.self_attn.q_proj.weight": "model-00003-of-00005.safetensors",
|
| 539 |
+
"decoder.layers.7.self_attn.k_proj.weight": "model-00003-of-00005.safetensors",
|
| 540 |
+
"decoder.layers.7.self_attn.v_proj.weight": "model-00003-of-00005.safetensors",
|
| 541 |
+
"decoder.layers.7.self_attn.q_norm.weight": "model-00003-of-00005.safetensors",
|
| 542 |
+
"decoder.layers.8.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
|
| 543 |
+
"decoder.layers.8.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
|
| 544 |
+
"decoder.layers.8.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
|
| 545 |
+
"decoder.layers.8.post_attention_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 546 |
+
"decoder.layers.8.post_feedforward_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 547 |
+
"decoder.layers.8.self_attn.k_norm.weight": "model-00003-of-00005.safetensors",
|
| 548 |
+
"decoder.layers.8.self_attn.o_proj.weight": "model-00003-of-00005.safetensors",
|
| 549 |
+
"decoder.layers.8.self_attn.q_proj.weight": "model-00003-of-00005.safetensors",
|
| 550 |
+
"decoder.layers.8.self_attn.k_proj.weight": "model-00003-of-00005.safetensors",
|
| 551 |
+
"decoder.layers.8.self_attn.v_proj.weight": "model-00003-of-00005.safetensors",
|
| 552 |
+
"decoder.layers.8.self_attn.q_norm.weight": "model-00003-of-00005.safetensors",
|
| 553 |
+
"decoder.layers.9.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
|
| 554 |
+
"decoder.layers.9.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
|
| 555 |
+
"decoder.layers.9.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
|
| 556 |
+
"decoder.layers.9.post_attention_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 557 |
+
"decoder.layers.9.post_feedforward_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 558 |
+
"decoder.layers.9.self_attn.k_norm.weight": "model-00003-of-00005.safetensors",
|
| 559 |
+
"decoder.layers.9.self_attn.o_proj.weight": "model-00003-of-00005.safetensors",
|
| 560 |
+
"decoder.layers.9.self_attn.q_proj.weight": "model-00003-of-00005.safetensors",
|
| 561 |
+
"decoder.layers.9.self_attn.k_proj.weight": "model-00003-of-00005.safetensors",
|
| 562 |
+
"decoder.layers.9.self_attn.v_proj.weight": "model-00003-of-00005.safetensors",
|
| 563 |
+
"decoder.layers.9.self_attn.q_norm.weight": "model-00003-of-00005.safetensors",
|
| 564 |
+
"decoder_read_concept_routes.0.residual_diag": "model-00003-of-00005.safetensors",
|
| 565 |
+
"decoder_read_concept_routes.0.w1.weight": "model-00003-of-00005.safetensors",
|
| 566 |
+
"decoder_read_concept_routes.0.w2.weight": "model-00003-of-00005.safetensors",
|
| 567 |
+
"decoder_read_concept_routes.1.residual_diag": "model-00003-of-00005.safetensors",
|
| 568 |
+
"decoder_read_concept_routes.1.w1.weight": "model-00003-of-00005.safetensors",
|
| 569 |
+
"decoder_read_concept_routes.1.w2.weight": "model-00003-of-00005.safetensors",
|
| 570 |
+
"decoder_read_concept_routes.10.residual_diag": "model-00003-of-00005.safetensors",
|
| 571 |
+
"decoder_read_concept_routes.10.w1.weight": "model-00003-of-00005.safetensors",
|
| 572 |
+
"decoder_read_concept_routes.10.w2.weight": "model-00003-of-00005.safetensors",
|
| 573 |
+
"decoder_read_concept_routes.11.residual_diag": "model-00003-of-00005.safetensors",
|
| 574 |
+
"decoder_read_concept_routes.11.w1.weight": "model-00003-of-00005.safetensors",
|
| 575 |
+
"decoder_read_concept_routes.11.w2.weight": "model-00003-of-00005.safetensors",
|
| 576 |
+
"decoder_read_concept_routes.12.residual_diag": "model-00003-of-00005.safetensors",
|
| 577 |
+
"decoder_read_concept_routes.12.w1.weight": "model-00003-of-00005.safetensors",
|
| 578 |
+
"decoder_read_concept_routes.12.w2.weight": "model-00003-of-00005.safetensors",
|
| 579 |
+
"decoder_read_concept_routes.13.residual_diag": "model-00003-of-00005.safetensors",
|
| 580 |
+
"decoder_read_concept_routes.13.w1.weight": "model-00003-of-00005.safetensors",
|
| 581 |
+
"decoder_read_concept_routes.13.w2.weight": "model-00003-of-00005.safetensors",
|
| 582 |
+
"decoder_read_concept_routes.14.residual_diag": "model-00003-of-00005.safetensors",
|
| 583 |
+
"decoder_read_concept_routes.14.w1.weight": "model-00003-of-00005.safetensors",
|
| 584 |
+
"decoder_read_concept_routes.14.w2.weight": "model-00003-of-00005.safetensors",
|
| 585 |
+
"decoder_read_concept_routes.15.residual_diag": "model-00003-of-00005.safetensors",
|
| 586 |
+
"decoder_read_concept_routes.15.w1.weight": "model-00003-of-00005.safetensors",
|
| 587 |
+
"decoder_read_concept_routes.15.w2.weight": "model-00003-of-00005.safetensors",
|
| 588 |
+
"decoder_read_concept_routes.2.residual_diag": "model-00003-of-00005.safetensors",
|
| 589 |
+
"decoder_read_concept_routes.2.w1.weight": "model-00003-of-00005.safetensors",
|
| 590 |
+
"decoder_read_concept_routes.2.w2.weight": "model-00003-of-00005.safetensors",
|
| 591 |
+
"decoder_read_concept_routes.3.residual_diag": "model-00003-of-00005.safetensors",
|
| 592 |
+
"decoder_read_concept_routes.3.w1.weight": "model-00003-of-00005.safetensors",
|
| 593 |
+
"decoder_read_concept_routes.3.w2.weight": "model-00003-of-00005.safetensors",
|
| 594 |
+
"decoder_read_concept_routes.4.residual_diag": "model-00003-of-00005.safetensors",
|
| 595 |
+
"decoder_read_concept_routes.4.w1.weight": "model-00003-of-00005.safetensors",
|
| 596 |
+
"decoder_read_concept_routes.4.w2.weight": "model-00003-of-00005.safetensors",
|
| 597 |
+
"decoder_read_concept_routes.5.residual_diag": "model-00003-of-00005.safetensors",
|
| 598 |
+
"decoder_read_concept_routes.5.w1.weight": "model-00003-of-00005.safetensors",
|
| 599 |
+
"decoder_read_concept_routes.5.w2.weight": "model-00003-of-00005.safetensors",
|
| 600 |
+
"decoder_read_concept_routes.6.residual_diag": "model-00003-of-00005.safetensors",
|
| 601 |
+
"decoder_read_concept_routes.6.w1.weight": "model-00003-of-00005.safetensors",
|
| 602 |
+
"decoder_read_concept_routes.6.w2.weight": "model-00003-of-00005.safetensors",
|
| 603 |
+
"decoder_read_concept_routes.7.residual_diag": "model-00003-of-00005.safetensors",
|
| 604 |
+
"decoder_read_concept_routes.7.w1.weight": "model-00003-of-00005.safetensors",
|
| 605 |
+
"decoder_read_concept_routes.7.w2.weight": "model-00003-of-00005.safetensors",
|
| 606 |
+
"decoder_read_concept_routes.8.residual_diag": "model-00003-of-00005.safetensors",
|
| 607 |
+
"decoder_read_concept_routes.8.w1.weight": "model-00003-of-00005.safetensors",
|
| 608 |
+
"decoder_read_concept_routes.8.w2.weight": "model-00003-of-00005.safetensors",
|
| 609 |
+
"decoder_read_concept_routes.9.residual_diag": "model-00003-of-00005.safetensors",
|
| 610 |
+
"decoder_read_concept_routes.9.w1.weight": "model-00003-of-00005.safetensors",
|
| 611 |
+
"decoder_read_concept_routes.9.w2.weight": "model-00003-of-00005.safetensors",
|
| 612 |
+
"decoder_read_concept_shared_source_norm.bias": "model-00003-of-00005.safetensors",
|
| 613 |
+
"decoder_read_concept_shared_source_norm.weight": "model-00003-of-00005.safetensors",
|
| 614 |
+
"decoder_read_encoder_routes.0.residual_diag": "model-00003-of-00005.safetensors",
|
| 615 |
+
"decoder_read_encoder_routes.0.w1.weight": "model-00003-of-00005.safetensors",
|
| 616 |
+
"decoder_read_encoder_routes.0.w2.weight": "model-00003-of-00005.safetensors",
|
| 617 |
+
"decoder_read_encoder_routes.1.residual_diag": "model-00003-of-00005.safetensors",
|
| 618 |
+
"decoder_read_encoder_routes.1.w1.weight": "model-00003-of-00005.safetensors",
|
| 619 |
+
"decoder_read_encoder_routes.1.w2.weight": "model-00003-of-00005.safetensors",
|
| 620 |
+
"decoder_read_encoder_routes.10.residual_diag": "model-00003-of-00005.safetensors",
|
| 621 |
+
"decoder_read_encoder_routes.10.w1.weight": "model-00003-of-00005.safetensors",
|
| 622 |
+
"decoder_read_encoder_routes.10.w2.weight": "model-00003-of-00005.safetensors",
|
| 623 |
+
"decoder_read_encoder_routes.11.residual_diag": "model-00003-of-00005.safetensors",
|
| 624 |
+
"decoder_read_encoder_routes.11.w1.weight": "model-00003-of-00005.safetensors",
|
| 625 |
+
"decoder_read_encoder_routes.11.w2.weight": "model-00003-of-00005.safetensors",
|
| 626 |
+
"decoder_read_encoder_routes.12.residual_diag": "model-00003-of-00005.safetensors",
|
| 627 |
+
"decoder_read_encoder_routes.12.w1.weight": "model-00003-of-00005.safetensors",
|
| 628 |
+
"decoder_read_encoder_routes.12.w2.weight": "model-00003-of-00005.safetensors",
|
| 629 |
+
"decoder_read_encoder_routes.13.residual_diag": "model-00003-of-00005.safetensors",
|
| 630 |
+
"decoder_read_encoder_routes.13.w1.weight": "model-00003-of-00005.safetensors",
|
| 631 |
+
"decoder_read_encoder_routes.13.w2.weight": "model-00003-of-00005.safetensors",
|
| 632 |
+
"decoder_read_encoder_routes.14.residual_diag": "model-00003-of-00005.safetensors",
|
| 633 |
+
"decoder_read_encoder_routes.14.w1.weight": "model-00003-of-00005.safetensors",
|
| 634 |
+
"decoder_read_encoder_routes.14.w2.weight": "model-00003-of-00005.safetensors",
|
| 635 |
+
"decoder_read_encoder_routes.15.residual_diag": "model-00003-of-00005.safetensors",
|
| 636 |
+
"decoder_read_encoder_routes.15.w1.weight": "model-00003-of-00005.safetensors",
|
| 637 |
+
"decoder_read_encoder_routes.15.w2.weight": "model-00003-of-00005.safetensors",
|
| 638 |
+
"decoder_read_encoder_routes.2.residual_diag": "model-00003-of-00005.safetensors",
|
| 639 |
+
"decoder_read_encoder_routes.2.w1.weight": "model-00003-of-00005.safetensors",
|
| 640 |
+
"decoder_read_encoder_routes.2.w2.weight": "model-00003-of-00005.safetensors",
|
| 641 |
+
"decoder_read_encoder_routes.3.residual_diag": "model-00003-of-00005.safetensors",
|
| 642 |
+
"decoder_read_encoder_routes.3.w1.weight": "model-00003-of-00005.safetensors",
|
| 643 |
+
"decoder_read_encoder_routes.3.w2.weight": "model-00003-of-00005.safetensors",
|
| 644 |
+
"decoder_read_encoder_routes.4.residual_diag": "model-00003-of-00005.safetensors",
|
| 645 |
+
"decoder_read_encoder_routes.4.w1.weight": "model-00003-of-00005.safetensors",
|
| 646 |
+
"decoder_read_encoder_routes.4.w2.weight": "model-00003-of-00005.safetensors",
|
| 647 |
+
"decoder_read_encoder_routes.5.residual_diag": "model-00003-of-00005.safetensors",
|
| 648 |
+
"decoder_read_encoder_routes.5.w1.weight": "model-00003-of-00005.safetensors",
|
| 649 |
+
"decoder_read_encoder_routes.5.w2.weight": "model-00003-of-00005.safetensors",
|
| 650 |
+
"decoder_read_encoder_routes.6.residual_diag": "model-00003-of-00005.safetensors",
|
| 651 |
+
"decoder_read_encoder_routes.6.w1.weight": "model-00003-of-00005.safetensors",
|
| 652 |
+
"decoder_read_encoder_routes.6.w2.weight": "model-00003-of-00005.safetensors",
|
| 653 |
+
"decoder_read_encoder_routes.7.residual_diag": "model-00003-of-00005.safetensors",
|
| 654 |
+
"decoder_read_encoder_routes.7.w1.weight": "model-00003-of-00005.safetensors",
|
| 655 |
+
"decoder_read_encoder_routes.7.w2.weight": "model-00003-of-00005.safetensors",
|
| 656 |
+
"decoder_read_encoder_routes.8.residual_diag": "model-00003-of-00005.safetensors",
|
| 657 |
+
"decoder_read_encoder_routes.8.w1.weight": "model-00003-of-00005.safetensors",
|
| 658 |
+
"decoder_read_encoder_routes.8.w2.weight": "model-00003-of-00005.safetensors",
|
| 659 |
+
"decoder_read_encoder_routes.9.residual_diag": "model-00003-of-00005.safetensors",
|
| 660 |
+
"decoder_read_encoder_routes.9.w1.weight": "model-00003-of-00005.safetensors",
|
| 661 |
+
"decoder_read_encoder_routes.9.w2.weight": "model-00003-of-00005.safetensors",
|
| 662 |
+
"decoder_read_encoder_shared_source_norm.bias": "model-00003-of-00005.safetensors",
|
| 663 |
+
"decoder_read_encoder_shared_source_norm.weight": "model-00003-of-00005.safetensors",
|
| 664 |
+
"embed_tokens.weight": "model-00003-of-00005.safetensors",
|
| 665 |
+
"encoder.layers.0.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
|
| 666 |
+
"encoder.layers.0.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
|
| 667 |
+
"encoder.layers.0.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
|
| 668 |
+
"encoder.layers.0.post_attention_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 669 |
+
"encoder.layers.0.post_feedforward_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 670 |
+
"encoder.layers.0.self_attn.k_norm.weight": "model-00003-of-00005.safetensors",
|
| 671 |
+
"encoder.layers.0.self_attn.o_proj.weight": "model-00003-of-00005.safetensors",
|
| 672 |
+
"encoder.layers.0.self_attn.q_proj.weight": "model-00003-of-00005.safetensors",
|
| 673 |
+
"encoder.layers.0.self_attn.k_proj.weight": "model-00003-of-00005.safetensors",
|
| 674 |
+
"encoder.layers.0.self_attn.v_proj.weight": "model-00003-of-00005.safetensors",
|
| 675 |
+
"encoder.layers.0.self_attn.q_norm.weight": "model-00003-of-00005.safetensors",
|
| 676 |
+
"encoder.layers.1.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
|
| 677 |
+
"encoder.layers.1.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
|
| 678 |
+
"encoder.layers.1.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
|
| 679 |
+
"encoder.layers.1.post_attention_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 680 |
+
"encoder.layers.1.post_feedforward_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 681 |
+
"encoder.layers.1.self_attn.k_norm.weight": "model-00003-of-00005.safetensors",
|
| 682 |
+
"encoder.layers.1.self_attn.o_proj.weight": "model-00003-of-00005.safetensors",
|
| 683 |
+
"encoder.layers.1.self_attn.q_proj.weight": "model-00003-of-00005.safetensors",
|
| 684 |
+
"encoder.layers.1.self_attn.k_proj.weight": "model-00003-of-00005.safetensors",
|
| 685 |
+
"encoder.layers.1.self_attn.v_proj.weight": "model-00003-of-00005.safetensors",
|
| 686 |
+
"encoder.layers.1.self_attn.q_norm.weight": "model-00003-of-00005.safetensors",
|
| 687 |
+
"encoder.layers.10.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
|
| 688 |
+
"encoder.layers.10.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
|
| 689 |
+
"encoder.layers.10.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
|
| 690 |
+
"encoder.layers.10.post_attention_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 691 |
+
"encoder.layers.10.post_feedforward_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 692 |
+
"encoder.layers.10.self_attn.k_norm.weight": "model-00003-of-00005.safetensors",
|
| 693 |
+
"encoder.layers.10.self_attn.o_proj.weight": "model-00003-of-00005.safetensors",
|
| 694 |
+
"encoder.layers.10.self_attn.q_proj.weight": "model-00003-of-00005.safetensors",
|
| 695 |
+
"encoder.layers.10.self_attn.k_proj.weight": "model-00003-of-00005.safetensors",
|
| 696 |
+
"encoder.layers.10.self_attn.v_proj.weight": "model-00003-of-00005.safetensors",
|
| 697 |
+
"encoder.layers.10.self_attn.q_norm.weight": "model-00003-of-00005.safetensors",
|
| 698 |
+
"encoder.layers.11.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
|
| 699 |
+
"encoder.layers.11.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
|
| 700 |
+
"encoder.layers.11.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
|
| 701 |
+
"encoder.layers.11.post_attention_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 702 |
+
"encoder.layers.11.post_feedforward_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 703 |
+
"encoder.layers.11.self_attn.k_norm.weight": "model-00003-of-00005.safetensors",
|
| 704 |
+
"encoder.layers.11.self_attn.o_proj.weight": "model-00003-of-00005.safetensors",
|
| 705 |
+
"encoder.layers.11.self_attn.q_proj.weight": "model-00003-of-00005.safetensors",
|
| 706 |
+
"encoder.layers.11.self_attn.k_proj.weight": "model-00003-of-00005.safetensors",
|
| 707 |
+
"encoder.layers.11.self_attn.v_proj.weight": "model-00003-of-00005.safetensors",
|
| 708 |
+
"encoder.layers.11.self_attn.q_norm.weight": "model-00003-of-00005.safetensors",
|
| 709 |
+
"encoder.layers.12.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
|
| 710 |
+
"encoder.layers.12.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
|
| 711 |
+
"encoder.layers.12.mlp.down_proj.weight": "model-00003-of-00005.safetensors",
|
| 712 |
+
"encoder.layers.12.post_attention_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 713 |
+
"encoder.layers.12.post_feedforward_layernorm.weight": "model-00003-of-00005.safetensors",
|
| 714 |
+
"encoder.layers.12.self_attn.k_norm.weight": "model-00003-of-00005.safetensors",
|
| 715 |
+
"encoder.layers.12.self_attn.o_proj.weight": "model-00003-of-00005.safetensors",
|
| 716 |
+
"encoder.layers.12.self_attn.q_proj.weight": "model-00003-of-00005.safetensors",
|
| 717 |
+
"encoder.layers.12.self_attn.k_proj.weight": "model-00003-of-00005.safetensors",
|
| 718 |
+
"encoder.layers.12.self_attn.v_proj.weight": "model-00003-of-00005.safetensors",
|
| 719 |
+
"encoder.layers.12.self_attn.q_norm.weight": "model-00003-of-00005.safetensors",
|
| 720 |
+
"encoder.layers.13.mlp.gate_proj.weight": "model-00003-of-00005.safetensors",
|
| 721 |
+
"encoder.layers.13.mlp.up_proj.weight": "model-00003-of-00005.safetensors",
|
| 722 |
+
"encoder.layers.13.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 723 |
+
"encoder.layers.13.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 724 |
+
"encoder.layers.13.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 725 |
+
"encoder.layers.13.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 726 |
+
"encoder.layers.13.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 727 |
+
"encoder.layers.13.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 728 |
+
"encoder.layers.13.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 729 |
+
"encoder.layers.13.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 730 |
+
"encoder.layers.13.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 731 |
+
"encoder.layers.14.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
|
| 732 |
+
"encoder.layers.14.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
|
| 733 |
+
"encoder.layers.14.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 734 |
+
"encoder.layers.14.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 735 |
+
"encoder.layers.14.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 736 |
+
"encoder.layers.14.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 737 |
+
"encoder.layers.14.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 738 |
+
"encoder.layers.14.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 739 |
+
"encoder.layers.14.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 740 |
+
"encoder.layers.14.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 741 |
+
"encoder.layers.14.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 742 |
+
"encoder.layers.15.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
|
| 743 |
+
"encoder.layers.15.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
|
| 744 |
+
"encoder.layers.15.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 745 |
+
"encoder.layers.15.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 746 |
+
"encoder.layers.15.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 747 |
+
"encoder.layers.15.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 748 |
+
"encoder.layers.15.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 749 |
+
"encoder.layers.15.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 750 |
+
"encoder.layers.15.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 751 |
+
"encoder.layers.15.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 752 |
+
"encoder.layers.15.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 753 |
+
"encoder.layers.2.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
|
| 754 |
+
"encoder.layers.2.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
|
| 755 |
+
"encoder.layers.2.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 756 |
+
"encoder.layers.2.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 757 |
+
"encoder.layers.2.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 758 |
+
"encoder.layers.2.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 759 |
+
"encoder.layers.2.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 760 |
+
"encoder.layers.2.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 761 |
+
"encoder.layers.2.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 762 |
+
"encoder.layers.2.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 763 |
+
"encoder.layers.2.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 764 |
+
"encoder.layers.3.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
|
| 765 |
+
"encoder.layers.3.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
|
| 766 |
+
"encoder.layers.3.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 767 |
+
"encoder.layers.3.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 768 |
+
"encoder.layers.3.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 769 |
+
"encoder.layers.3.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 770 |
+
"encoder.layers.3.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 771 |
+
"encoder.layers.3.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 772 |
+
"encoder.layers.3.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 773 |
+
"encoder.layers.3.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 774 |
+
"encoder.layers.3.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 775 |
+
"encoder.layers.4.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
|
| 776 |
+
"encoder.layers.4.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
|
| 777 |
+
"encoder.layers.4.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 778 |
+
"encoder.layers.4.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 779 |
+
"encoder.layers.4.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 780 |
+
"encoder.layers.4.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 781 |
+
"encoder.layers.4.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 782 |
+
"encoder.layers.4.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 783 |
+
"encoder.layers.4.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 784 |
+
"encoder.layers.4.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 785 |
+
"encoder.layers.4.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 786 |
+
"encoder.layers.5.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
|
| 787 |
+
"encoder.layers.5.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
|
| 788 |
+
"encoder.layers.5.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 789 |
+
"encoder.layers.5.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 790 |
+
"encoder.layers.5.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 791 |
+
"encoder.layers.5.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 792 |
+
"encoder.layers.5.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 793 |
+
"encoder.layers.5.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 794 |
+
"encoder.layers.5.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 795 |
+
"encoder.layers.5.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 796 |
+
"encoder.layers.5.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 797 |
+
"encoder.layers.6.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
|
| 798 |
+
"encoder.layers.6.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
|
| 799 |
+
"encoder.layers.6.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 800 |
+
"encoder.layers.6.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 801 |
+
"encoder.layers.6.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 802 |
+
"encoder.layers.6.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 803 |
+
"encoder.layers.6.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 804 |
+
"encoder.layers.6.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 805 |
+
"encoder.layers.6.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 806 |
+
"encoder.layers.6.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 807 |
+
"encoder.layers.6.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 808 |
+
"encoder.layers.7.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
|
| 809 |
+
"encoder.layers.7.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
|
| 810 |
+
"encoder.layers.7.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 811 |
+
"encoder.layers.7.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 812 |
+
"encoder.layers.7.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 813 |
+
"encoder.layers.7.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 814 |
+
"encoder.layers.7.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 815 |
+
"encoder.layers.7.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 816 |
+
"encoder.layers.7.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 817 |
+
"encoder.layers.7.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 818 |
+
"encoder.layers.7.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 819 |
+
"encoder.layers.8.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
|
| 820 |
+
"encoder.layers.8.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
|
| 821 |
+
"encoder.layers.8.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 822 |
+
"encoder.layers.8.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 823 |
+
"encoder.layers.8.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 824 |
+
"encoder.layers.8.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 825 |
+
"encoder.layers.8.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 826 |
+
"encoder.layers.8.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 827 |
+
"encoder.layers.8.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 828 |
+
"encoder.layers.8.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 829 |
+
"encoder.layers.8.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 830 |
+
"encoder.layers.9.mlp.gate_proj.weight": "model-00004-of-00005.safetensors",
|
| 831 |
+
"encoder.layers.9.mlp.up_proj.weight": "model-00004-of-00005.safetensors",
|
| 832 |
+
"encoder.layers.9.mlp.down_proj.weight": "model-00004-of-00005.safetensors",
|
| 833 |
+
"encoder.layers.9.post_attention_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 834 |
+
"encoder.layers.9.post_feedforward_layernorm.weight": "model-00004-of-00005.safetensors",
|
| 835 |
+
"encoder.layers.9.self_attn.k_norm.weight": "model-00004-of-00005.safetensors",
|
| 836 |
+
"encoder.layers.9.self_attn.o_proj.weight": "model-00004-of-00005.safetensors",
|
| 837 |
+
"encoder.layers.9.self_attn.q_proj.weight": "model-00004-of-00005.safetensors",
|
| 838 |
+
"encoder.layers.9.self_attn.k_proj.weight": "model-00004-of-00005.safetensors",
|
| 839 |
+
"encoder.layers.9.self_attn.v_proj.weight": "model-00004-of-00005.safetensors",
|
| 840 |
+
"encoder.layers.9.self_attn.q_norm.weight": "model-00004-of-00005.safetensors",
|
| 841 |
+
"final_read_concept_gate_logits": "model-00004-of-00005.safetensors",
|
| 842 |
+
"fusion_hl_norm.bias": "model-00004-of-00005.safetensors",
|
| 843 |
+
"fusion_hl_norm.weight": "model-00004-of-00005.safetensors",
|
| 844 |
+
"fusion_norm_alpha": "model-00004-of-00005.safetensors",
|
| 845 |
+
"fusion_tok_norm.bias": "model-00004-of-00005.safetensors",
|
| 846 |
+
"fusion_tok_norm.weight": "model-00004-of-00005.safetensors",
|
| 847 |
+
"lm_head.weight": "model-00005-of-00005.safetensors"
|
| 848 |
+
}
|
| 849 |
+
}
|
modeling_ncp_olmo3.py
ADDED
|
@@ -0,0 +1,1232 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Standalone PyTorch/Transformers implementation of NCP OLMo3.
|
| 2 |
+
|
| 3 |
+
The module uses conventional Hugging Face projection names. The converter
|
| 4 |
+
losslessly splits the fused Megatron tensors once, so the exported SafeTensors
|
| 5 |
+
load without importing ConceptLM, Megatron, or vLLM and are shared with the
|
| 6 |
+
ConceptLM vLLM backend.
|
| 7 |
+
|
| 8 |
+
The first correctness milestone uses batch-1 full-prefix recomputation.
|
| 9 |
+
This keeps the inference graph small and auditable while numerical parity is
|
| 10 |
+
established. A native Transformers cache can be added independently after
|
| 11 |
+
the full-prefix path passes the reference gates.
|
| 12 |
+
"""
|
| 13 |
+
|
| 14 |
+
from __future__ import annotations
|
| 15 |
+
|
| 16 |
+
import json
|
| 17 |
+
import math
|
| 18 |
+
import os
|
| 19 |
+
from typing import Any, Optional
|
| 20 |
+
|
| 21 |
+
import torch
|
| 22 |
+
import torch.nn as nn
|
| 23 |
+
import torch.nn.functional as F
|
| 24 |
+
from transformers import PreTrainedModel
|
| 25 |
+
from transformers.generation import GenerationMixin
|
| 26 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast
|
| 27 |
+
|
| 28 |
+
from .configuration_ncp_olmo3 import NCPOlmo3Config
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
class NCPRMSNorm(nn.Module):
|
| 32 |
+
def __init__(self, hidden_size: int, eps: float) -> None:
|
| 33 |
+
super().__init__()
|
| 34 |
+
self.weight = nn.Parameter(torch.empty(hidden_size))
|
| 35 |
+
self.eps = float(eps)
|
| 36 |
+
|
| 37 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 38 |
+
input_dtype = hidden_states.dtype
|
| 39 |
+
hidden_states = hidden_states.float()
|
| 40 |
+
variance = hidden_states.pow(2).mean(dim=-1, keepdim=True)
|
| 41 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
|
| 42 |
+
return (self.weight * hidden_states).to(input_dtype)
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def _rotate_half_neox(x: torch.Tensor) -> torch.Tensor:
|
| 46 |
+
first, second = x.chunk(2, dim=-1)
|
| 47 |
+
return torch.cat((-second, first), dim=-1)
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
def _yarn_find_correction_dim(
|
| 51 |
+
num_rotations: float, dim: int, theta: float, original_max_position_embeddings: int
|
| 52 |
+
) -> float:
|
| 53 |
+
return (
|
| 54 |
+
dim
|
| 55 |
+
* math.log(original_max_position_embeddings / (num_rotations * 2.0 * math.pi))
|
| 56 |
+
/ (2.0 * math.log(theta))
|
| 57 |
+
)
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
def _yarn_find_correction_range(
|
| 61 |
+
beta_fast: float,
|
| 62 |
+
beta_slow: float,
|
| 63 |
+
dim: int,
|
| 64 |
+
theta: float,
|
| 65 |
+
original_max_position_embeddings: int,
|
| 66 |
+
round_to_int: bool,
|
| 67 |
+
) -> tuple[float, float]:
|
| 68 |
+
low = _yarn_find_correction_dim(beta_fast, dim, theta, original_max_position_embeddings)
|
| 69 |
+
high = _yarn_find_correction_dim(beta_slow, dim, theta, original_max_position_embeddings)
|
| 70 |
+
if round_to_int:
|
| 71 |
+
low = float(math.floor(low))
|
| 72 |
+
high = float(math.ceil(high))
|
| 73 |
+
return max(low, 0.0), min(high, float(dim - 1))
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
def _yarn_linear_ramp_mask(low: float, high: float, dim: int, device: torch.device) -> torch.Tensor:
|
| 77 |
+
if low == high:
|
| 78 |
+
high += 0.001
|
| 79 |
+
ramp = (torch.arange(dim, dtype=torch.float32, device=device) - low) / (high - low)
|
| 80 |
+
return ramp.clamp(0.0, 1.0)
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
def _yarn_mscale(scale: float, multiplier: float) -> float:
|
| 84 |
+
if scale <= 1.0:
|
| 85 |
+
return 1.0
|
| 86 |
+
return 0.1 * multiplier * math.log(scale) + 1.0
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
class NCPRotaryEmbedding(nn.Module):
|
| 90 |
+
def __init__(
|
| 91 |
+
self,
|
| 92 |
+
head_dim: int,
|
| 93 |
+
theta: float,
|
| 94 |
+
*,
|
| 95 |
+
apply_yarn: bool = False,
|
| 96 |
+
scaling_factor: float = 1.0,
|
| 97 |
+
original_max_position_embeddings: int = 4096,
|
| 98 |
+
beta_fast: float = 32.0,
|
| 99 |
+
beta_slow: float = 1.0,
|
| 100 |
+
mscale: float = 1.0,
|
| 101 |
+
mscale_all_dim: float = 0.0,
|
| 102 |
+
correction_range_round_to_int: bool = True,
|
| 103 |
+
) -> None:
|
| 104 |
+
super().__init__()
|
| 105 |
+
self.head_dim = int(head_dim)
|
| 106 |
+
self.theta = float(theta)
|
| 107 |
+
self.apply_yarn = bool(apply_yarn)
|
| 108 |
+
self.scaling_factor = float(scaling_factor)
|
| 109 |
+
self.original_max_position_embeddings = int(original_max_position_embeddings)
|
| 110 |
+
self.beta_fast = float(beta_fast)
|
| 111 |
+
self.beta_slow = float(beta_slow)
|
| 112 |
+
self.mscale = float(mscale)
|
| 113 |
+
self.mscale_all_dim = float(mscale_all_dim)
|
| 114 |
+
self.correction_range_round_to_int = bool(correction_range_round_to_int)
|
| 115 |
+
|
| 116 |
+
def _build_inv_freq(self, device: torch.device) -> tuple[torch.Tensor, float]:
|
| 117 |
+
# Do not register this derived value as a non-persistent buffer.
|
| 118 |
+
# Transformers constructs remote-code models under a meta-device
|
| 119 |
+
# context in from_pretrained(); a buffer omitted from the checkpoint
|
| 120 |
+
# can otherwise be materialized without its initialized values.
|
| 121 |
+
inv_freq_extra = 1.0 / (
|
| 122 |
+
self.theta
|
| 123 |
+
** (
|
| 124 |
+
torch.arange(0, self.head_dim, 2, dtype=torch.float32, device=device)
|
| 125 |
+
/ float(self.head_dim)
|
| 126 |
+
)
|
| 127 |
+
)
|
| 128 |
+
if not self.apply_yarn:
|
| 129 |
+
return inv_freq_extra, 1.0
|
| 130 |
+
|
| 131 |
+
inv_freq_inter = inv_freq_extra / self.scaling_factor
|
| 132 |
+
low, high = _yarn_find_correction_range(
|
| 133 |
+
self.beta_fast,
|
| 134 |
+
self.beta_slow,
|
| 135 |
+
self.head_dim,
|
| 136 |
+
self.theta,
|
| 137 |
+
self.original_max_position_embeddings,
|
| 138 |
+
self.correction_range_round_to_int,
|
| 139 |
+
)
|
| 140 |
+
inv_freq_mask = 1.0 - _yarn_linear_ramp_mask(low, high, self.head_dim // 2, device)
|
| 141 |
+
inv_freq = inv_freq_inter * (1.0 - inv_freq_mask) + inv_freq_extra * inv_freq_mask
|
| 142 |
+
concentration = _yarn_mscale(self.scaling_factor, self.mscale) / _yarn_mscale(
|
| 143 |
+
self.scaling_factor, self.mscale_all_dim
|
| 144 |
+
)
|
| 145 |
+
return inv_freq, concentration
|
| 146 |
+
|
| 147 |
+
def forward(
|
| 148 |
+
self, query: torch.Tensor, key: torch.Tensor, position_ids: torch.Tensor
|
| 149 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 150 |
+
inv_freq, concentration = self._build_inv_freq(query.device)
|
| 151 |
+
# Match Megatron's FP32 `torch.outer(position, inv_freq)` semantics.
|
| 152 |
+
# An einsum here may lower to a TF32 GEMM on CUDA; its phase error is
|
| 153 |
+
# small at short positions but becomes material at long context.
|
| 154 |
+
frequencies = position_ids.float().unsqueeze(-1) * inv_freq.view(1, 1, -1)
|
| 155 |
+
embedding = torch.cat((frequencies, frequencies), dim=-1)
|
| 156 |
+
cos = (embedding.cos() * concentration).to(query.dtype).unsqueeze(1)
|
| 157 |
+
sin = (embedding.sin() * concentration).to(query.dtype).unsqueeze(1)
|
| 158 |
+
rotated_query = _rotate_half_neox(query)
|
| 159 |
+
rotated_key = _rotate_half_neox(key)
|
| 160 |
+
query_output = query * cos + rotated_query * sin
|
| 161 |
+
key_output = key * cos + rotated_key * sin
|
| 162 |
+
if os.environ.get("NCP_STANDALONE_DEBUG_FINITE") == "1" and (
|
| 163 |
+
not bool(torch.isfinite(query_output).all())
|
| 164 |
+
or not bool(torch.isfinite(key_output).all())
|
| 165 |
+
):
|
| 166 |
+
|
| 167 |
+
def summary(tensor: torch.Tensor) -> dict[str, Any]:
|
| 168 |
+
finite = torch.isfinite(tensor)
|
| 169 |
+
values = tensor[finite].float()
|
| 170 |
+
return {
|
| 171 |
+
"shape": list(tensor.shape),
|
| 172 |
+
"dtype": str(tensor.dtype),
|
| 173 |
+
"finite_fraction": float(finite.float().mean().item()),
|
| 174 |
+
"nan_count": int(torch.isnan(tensor).sum().item()),
|
| 175 |
+
"inf_count": int(torch.isinf(tensor).sum().item()),
|
| 176 |
+
"finite_min": (float(values.min().item()) if values.numel() else None),
|
| 177 |
+
"finite_max": (float(values.max().item()) if values.numel() else None),
|
| 178 |
+
}
|
| 179 |
+
|
| 180 |
+
raise FloatingPointError(
|
| 181 |
+
"NCP_STANDALONE_ROTARY_NONFINITE "
|
| 182 |
+
+ json.dumps(
|
| 183 |
+
{
|
| 184 |
+
"position_ids": position_ids.detach().cpu().tolist(),
|
| 185 |
+
"query": summary(query),
|
| 186 |
+
"key": summary(key),
|
| 187 |
+
"inv_freq": summary(inv_freq),
|
| 188 |
+
"cos": summary(cos),
|
| 189 |
+
"sin": summary(sin),
|
| 190 |
+
"rotated_query": summary(rotated_query),
|
| 191 |
+
"rotated_key": summary(rotated_key),
|
| 192 |
+
"query_output": summary(query_output),
|
| 193 |
+
"key_output": summary(key_output),
|
| 194 |
+
},
|
| 195 |
+
sort_keys=True,
|
| 196 |
+
)
|
| 197 |
+
)
|
| 198 |
+
query = query_output
|
| 199 |
+
key = key_output
|
| 200 |
+
return query, key
|
| 201 |
+
|
| 202 |
+
|
| 203 |
+
def _causal_mask(
|
| 204 |
+
position_ids: torch.Tensor,
|
| 205 |
+
key_length: int,
|
| 206 |
+
*,
|
| 207 |
+
device: torch.device,
|
| 208 |
+
sliding_window: Optional[int],
|
| 209 |
+
) -> torch.Tensor:
|
| 210 |
+
if position_ids.ndim != 2:
|
| 211 |
+
raise ValueError("position_ids must have [batch, sequence] shape")
|
| 212 |
+
query_positions = position_ids.to(device=device, dtype=torch.long)
|
| 213 |
+
key_start = query_positions[:, -1:] + 1 - int(key_length)
|
| 214 |
+
key_positions = key_start.unsqueeze(-1) + torch.arange(
|
| 215 |
+
int(key_length), device=device, dtype=torch.long
|
| 216 |
+
).view(1, 1, -1)
|
| 217 |
+
mask = key_positions <= query_positions.unsqueeze(-1)
|
| 218 |
+
if sliding_window is not None:
|
| 219 |
+
mask &= key_positions >= (query_positions.unsqueeze(-1) - int(sliding_window))
|
| 220 |
+
return mask
|
| 221 |
+
|
| 222 |
+
|
| 223 |
+
class NCPAttention(nn.Module):
|
| 224 |
+
def __init__(self, config: NCPOlmo3Config, layer_number: int) -> None:
|
| 225 |
+
super().__init__()
|
| 226 |
+
hidden_size = int(config.hidden_size)
|
| 227 |
+
num_heads = int(config.num_attention_heads)
|
| 228 |
+
if hidden_size % num_heads:
|
| 229 |
+
raise ValueError("hidden_size must be divisible by num_attention_heads")
|
| 230 |
+
self.hidden_size = hidden_size
|
| 231 |
+
self.num_heads = num_heads
|
| 232 |
+
self.head_dim = hidden_size // num_heads
|
| 233 |
+
# Keep the serialized parameter names and tensor boundaries aligned with
|
| 234 |
+
# the conventional Hugging Face attention contract. The source
|
| 235 |
+
# Megatron checkpoint stores Q/K/V interleaved in one tensor; the
|
| 236 |
+
# converter deinterleaves it once and writes these three projections.
|
| 237 |
+
self.q_proj = nn.Linear(hidden_size, hidden_size, bias=False)
|
| 238 |
+
self.k_proj = nn.Linear(hidden_size, hidden_size, bias=False)
|
| 239 |
+
self.v_proj = nn.Linear(hidden_size, hidden_size, bias=False)
|
| 240 |
+
eps = float(config.layernorm_epsilon)
|
| 241 |
+
self.qk_norm_mode = str(getattr(config, "qk_norm_mode", "full_hidden"))
|
| 242 |
+
if self.qk_norm_mode not in {"full_hidden", "per_head"}:
|
| 243 |
+
raise ValueError(
|
| 244 |
+
"qk_norm_mode must be 'full_hidden' or 'per_head', got "
|
| 245 |
+
f"{self.qk_norm_mode!r}"
|
| 246 |
+
)
|
| 247 |
+
qk_norm_size = self.head_dim if self.qk_norm_mode == "per_head" else hidden_size
|
| 248 |
+
self.q_norm = NCPRMSNorm(qk_norm_size, eps)
|
| 249 |
+
self.k_norm = NCPRMSNorm(qk_norm_size, eps)
|
| 250 |
+
skip_frequency = int(config.window_attn_skip_freq)
|
| 251 |
+
self.sliding_window = (
|
| 252 |
+
None if int(layer_number) % skip_frequency == 0 else int(config.window_size[0])
|
| 253 |
+
)
|
| 254 |
+
position_embedding_type = str(getattr(config, "position_embedding_type", "rope"))
|
| 255 |
+
apply_yarn = position_embedding_type == "yarn" and (
|
| 256 |
+
self.sliding_window is None
|
| 257 |
+
or not bool(getattr(config, "yarn_full_attn_layers_only", False))
|
| 258 |
+
)
|
| 259 |
+
if position_embedding_type not in {"rope", "yarn"}:
|
| 260 |
+
raise ValueError(
|
| 261 |
+
"standalone NCP OLMo3 supports rope or yarn position embeddings, "
|
| 262 |
+
f"got {position_embedding_type!r}"
|
| 263 |
+
)
|
| 264 |
+
self.rotary_emb = NCPRotaryEmbedding(
|
| 265 |
+
self.head_dim,
|
| 266 |
+
float(config.rotary_base),
|
| 267 |
+
apply_yarn=apply_yarn,
|
| 268 |
+
scaling_factor=float(getattr(config, "yarn_rotary_scaling_factor", 1.0)),
|
| 269 |
+
original_max_position_embeddings=int(
|
| 270 |
+
getattr(
|
| 271 |
+
config,
|
| 272 |
+
"yarn_original_max_position_embeddings",
|
| 273 |
+
getattr(config, "max_position_embeddings", 4096),
|
| 274 |
+
)
|
| 275 |
+
),
|
| 276 |
+
beta_fast=float(getattr(config, "yarn_beta_fast", 32.0)),
|
| 277 |
+
beta_slow=float(getattr(config, "yarn_beta_slow", 1.0)),
|
| 278 |
+
mscale=float(getattr(config, "yarn_mscale", 1.0)),
|
| 279 |
+
mscale_all_dim=float(getattr(config, "yarn_mscale_all_dim", 0.0)),
|
| 280 |
+
correction_range_round_to_int=bool(
|
| 281 |
+
getattr(config, "yarn_correction_range_round_to_int", True)
|
| 282 |
+
),
|
| 283 |
+
)
|
| 284 |
+
self.o_proj = nn.Linear(hidden_size, hidden_size, bias=False)
|
| 285 |
+
|
| 286 |
+
def forward(
|
| 287 |
+
self,
|
| 288 |
+
hidden_states: torch.Tensor,
|
| 289 |
+
position_ids: torch.Tensor,
|
| 290 |
+
attention_mask: Optional[torch.Tensor],
|
| 291 |
+
past_key_value: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
|
| 292 |
+
use_cache: bool = False,
|
| 293 |
+
) -> torch.Tensor | tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
|
| 294 |
+
batch_size, sequence_length, _ = hidden_states.shape
|
| 295 |
+
query = self.q_proj(hidden_states)
|
| 296 |
+
key = self.k_proj(hidden_states)
|
| 297 |
+
value = (
|
| 298 |
+
self.v_proj(hidden_states)
|
| 299 |
+
.view(batch_size, sequence_length, self.num_heads, self.head_dim)
|
| 300 |
+
.transpose(1, 2)
|
| 301 |
+
)
|
| 302 |
+
if self.qk_norm_mode == "per_head":
|
| 303 |
+
query = self.q_norm(
|
| 304 |
+
query.view(batch_size, sequence_length, self.num_heads, self.head_dim)
|
| 305 |
+
).transpose(1, 2)
|
| 306 |
+
key = self.k_norm(
|
| 307 |
+
key.view(batch_size, sequence_length, self.num_heads, self.head_dim)
|
| 308 |
+
).transpose(1, 2)
|
| 309 |
+
else:
|
| 310 |
+
query = (
|
| 311 |
+
self.q_norm(query)
|
| 312 |
+
.view(batch_size, sequence_length, self.num_heads, self.head_dim)
|
| 313 |
+
.transpose(1, 2)
|
| 314 |
+
)
|
| 315 |
+
key = (
|
| 316 |
+
self.k_norm(key)
|
| 317 |
+
.view(batch_size, sequence_length, self.num_heads, self.head_dim)
|
| 318 |
+
.transpose(1, 2)
|
| 319 |
+
)
|
| 320 |
+
query, key = self.rotary_emb(query, key, position_ids)
|
| 321 |
+
|
| 322 |
+
if past_key_value is not None:
|
| 323 |
+
past_key, past_value = past_key_value
|
| 324 |
+
key = torch.cat((past_key, key), dim=2)
|
| 325 |
+
value = torch.cat((past_value, value), dim=2)
|
| 326 |
+
key_length = int(key.shape[2])
|
| 327 |
+
mask = _causal_mask(
|
| 328 |
+
position_ids,
|
| 329 |
+
key_length,
|
| 330 |
+
device=hidden_states.device,
|
| 331 |
+
sliding_window=self.sliding_window,
|
| 332 |
+
)
|
| 333 |
+
mask = mask.unsqueeze(1)
|
| 334 |
+
if attention_mask is not None:
|
| 335 |
+
if attention_mask.ndim != 2:
|
| 336 |
+
raise ValueError("attention_mask must have [batch, sequence] shape")
|
| 337 |
+
if int(attention_mask.shape[1]) < key_length:
|
| 338 |
+
raise ValueError("attention_mask is shorter than the attention keys")
|
| 339 |
+
key_attention_mask = attention_mask[:, -key_length:].bool()
|
| 340 |
+
mask = mask & key_attention_mask[:, None, None, :]
|
| 341 |
+
output = F.scaled_dot_product_attention(
|
| 342 |
+
query, key, value, attn_mask=mask, dropout_p=0.0, scale=self.head_dim**-0.5
|
| 343 |
+
)
|
| 344 |
+
output = output.transpose(1, 2).reshape(batch_size, sequence_length, self.hidden_size)
|
| 345 |
+
output = self.o_proj(output)
|
| 346 |
+
if not use_cache:
|
| 347 |
+
return output
|
| 348 |
+
if self.sliding_window is not None:
|
| 349 |
+
retained = int(self.sliding_window) + 1
|
| 350 |
+
key = key[:, :, -retained:, :]
|
| 351 |
+
value = value[:, :, -retained:, :]
|
| 352 |
+
return output, (key, value)
|
| 353 |
+
|
| 354 |
+
|
| 355 |
+
class NCPMLP(nn.Module):
|
| 356 |
+
def __init__(self, config: NCPOlmo3Config) -> None:
|
| 357 |
+
super().__init__()
|
| 358 |
+
hidden_size = int(config.hidden_size)
|
| 359 |
+
intermediate_size = int(config.ffn_hidden_size)
|
| 360 |
+
self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
|
| 361 |
+
self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
|
| 362 |
+
self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
|
| 363 |
+
|
| 364 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 365 |
+
gate = self.gate_proj(hidden_states)
|
| 366 |
+
up = self.up_proj(hidden_states)
|
| 367 |
+
return self.down_proj(F.silu(gate) * up)
|
| 368 |
+
|
| 369 |
+
|
| 370 |
+
class NCPOlmo3Layer(nn.Module):
|
| 371 |
+
def __init__(self, config: NCPOlmo3Config, layer_number: int) -> None:
|
| 372 |
+
super().__init__()
|
| 373 |
+
self.self_attn = NCPAttention(config, layer_number)
|
| 374 |
+
self.mlp = NCPMLP(config)
|
| 375 |
+
eps = float(config.layernorm_epsilon)
|
| 376 |
+
self.post_attention_layernorm = NCPRMSNorm(int(config.hidden_size), eps)
|
| 377 |
+
self.post_feedforward_layernorm = NCPRMSNorm(int(config.hidden_size), eps)
|
| 378 |
+
|
| 379 |
+
def forward(
|
| 380 |
+
self,
|
| 381 |
+
hidden_states: torch.Tensor,
|
| 382 |
+
position_ids: torch.Tensor,
|
| 383 |
+
attention_mask: Optional[torch.Tensor],
|
| 384 |
+
past_key_value: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
|
| 385 |
+
use_cache: bool = False,
|
| 386 |
+
) -> torch.Tensor | tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
|
| 387 |
+
residual = hidden_states
|
| 388 |
+
attention_output = self.self_attn(
|
| 389 |
+
hidden_states,
|
| 390 |
+
position_ids,
|
| 391 |
+
attention_mask,
|
| 392 |
+
past_key_value=past_key_value,
|
| 393 |
+
use_cache=use_cache,
|
| 394 |
+
)
|
| 395 |
+
if use_cache:
|
| 396 |
+
hidden_states, present_key_value = attention_output
|
| 397 |
+
else:
|
| 398 |
+
hidden_states = attention_output
|
| 399 |
+
present_key_value = None
|
| 400 |
+
hidden_states = residual + self.post_attention_layernorm(hidden_states)
|
| 401 |
+
residual = hidden_states
|
| 402 |
+
hidden_states = self.mlp(hidden_states)
|
| 403 |
+
hidden_states = residual + self.post_feedforward_layernorm(hidden_states)
|
| 404 |
+
if use_cache:
|
| 405 |
+
if present_key_value is None:
|
| 406 |
+
raise RuntimeError("attention did not return a cache")
|
| 407 |
+
return hidden_states, present_key_value
|
| 408 |
+
return hidden_states
|
| 409 |
+
|
| 410 |
+
|
| 411 |
+
class NCPTower(nn.Module):
|
| 412 |
+
def __init__(self, config: NCPOlmo3Config, num_layers: int, *, final_layernorm: bool) -> None:
|
| 413 |
+
super().__init__()
|
| 414 |
+
self.layers = nn.ModuleList(
|
| 415 |
+
[NCPOlmo3Layer(config, layer_index + 1) for layer_index in range(int(num_layers))]
|
| 416 |
+
)
|
| 417 |
+
self.norm = (
|
| 418 |
+
NCPRMSNorm(int(config.hidden_size), float(config.layernorm_epsilon))
|
| 419 |
+
if final_layernorm
|
| 420 |
+
else None
|
| 421 |
+
)
|
| 422 |
+
|
| 423 |
+
|
| 424 |
+
class NCPDepthDD(nn.Module):
|
| 425 |
+
def __init__(self, hidden_size: int, layer_index: int, *, use_softmax: bool = False) -> None:
|
| 426 |
+
super().__init__()
|
| 427 |
+
num_previous = int(layer_index) + 2
|
| 428 |
+
self.use_softmax = bool(use_softmax)
|
| 429 |
+
self.static_a = nn.Parameter(torch.empty(num_previous))
|
| 430 |
+
self.w1 = nn.Linear(hidden_size, num_previous, bias=False)
|
| 431 |
+
self.w2 = nn.Linear(num_previous, num_previous, bias=False)
|
| 432 |
+
|
| 433 |
+
def forward(
|
| 434 |
+
self, current_hidden: torch.Tensor, history_states: list[torch.Tensor]
|
| 435 |
+
) -> torch.Tensor:
|
| 436 |
+
route_weights = self.w2(F.gelu(self.w1(current_hidden)))
|
| 437 |
+
route_weights = route_weights + self.static_a
|
| 438 |
+
if self.use_softmax:
|
| 439 |
+
route_weights = route_weights.softmax(dim=-1)
|
| 440 |
+
history = torch.stack(history_states, dim=-2)
|
| 441 |
+
return (route_weights.unsqueeze(-1) * history).sum(dim=-2)
|
| 442 |
+
|
| 443 |
+
|
| 444 |
+
class NCPSelfDD(nn.Module):
|
| 445 |
+
def __init__(self, hidden_size: int, num_layers: int) -> None:
|
| 446 |
+
super().__init__()
|
| 447 |
+
self.depth_dds = nn.ModuleList(
|
| 448 |
+
[NCPDepthDD(hidden_size, layer_index) for layer_index in range(int(num_layers))]
|
| 449 |
+
)
|
| 450 |
+
|
| 451 |
+
|
| 452 |
+
class NCPSelfCumsumDD(nn.Module):
|
| 453 |
+
"""Single-state depth recurrence used by the trained cumsum checkpoints."""
|
| 454 |
+
|
| 455 |
+
def __init__(self) -> None:
|
| 456 |
+
super().__init__()
|
| 457 |
+
self.alpha = nn.Parameter(torch.empty(()))
|
| 458 |
+
|
| 459 |
+
def initial_state(self, module_input: torch.Tensor) -> torch.Tensor:
|
| 460 |
+
return module_input
|
| 461 |
+
|
| 462 |
+
def forward(
|
| 463 |
+
self,
|
| 464 |
+
current_hidden: torch.Tensor,
|
| 465 |
+
previous_state: torch.Tensor,
|
| 466 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 467 |
+
alpha = torch.tanh(self.alpha).to(
|
| 468 |
+
dtype=current_hidden.dtype,
|
| 469 |
+
device=current_hidden.device,
|
| 470 |
+
)
|
| 471 |
+
next_state = current_hidden + alpha * previous_state
|
| 472 |
+
return next_state, next_state
|
| 473 |
+
|
| 474 |
+
|
| 475 |
+
class NCPDiagResidualRoute(nn.Module):
|
| 476 |
+
def __init__(self, hidden_size: int, num_sources: int) -> None:
|
| 477 |
+
super().__init__()
|
| 478 |
+
self.residual_diag = nn.Parameter(torch.empty(hidden_size))
|
| 479 |
+
self.w1 = nn.Linear(hidden_size, num_sources, bias=False)
|
| 480 |
+
self.w2 = nn.Linear(num_sources, num_sources, bias=False)
|
| 481 |
+
|
| 482 |
+
def forward(
|
| 483 |
+
self,
|
| 484 |
+
target_hidden: torch.Tensor,
|
| 485 |
+
normalized_sources: list[torch.Tensor],
|
| 486 |
+
residual_scale: Optional[torch.Tensor] = None,
|
| 487 |
+
) -> torch.Tensor:
|
| 488 |
+
weights = self.w2(F.gelu(self.w1(target_hidden))).softmax(dim=-1)
|
| 489 |
+
sources = torch.stack(normalized_sources, dim=-2)
|
| 490 |
+
source_mix = (weights.unsqueeze(-1) * sources).sum(dim=-2)
|
| 491 |
+
update = source_mix * self.residual_diag.to(source_mix.dtype)
|
| 492 |
+
if residual_scale is not None:
|
| 493 |
+
update = update * residual_scale.to(update.dtype)
|
| 494 |
+
return target_hidden + update
|
| 495 |
+
|
| 496 |
+
|
| 497 |
+
class NCPCrossCumsumRoute(nn.Module):
|
| 498 |
+
"""Read the final normalized source state with the trained scalar beta."""
|
| 499 |
+
|
| 500 |
+
def __init__(self) -> None:
|
| 501 |
+
super().__init__()
|
| 502 |
+
self.beta = nn.Parameter(torch.empty(()))
|
| 503 |
+
|
| 504 |
+
def forward(
|
| 505 |
+
self,
|
| 506 |
+
target_hidden: torch.Tensor,
|
| 507 |
+
normalized_sources: list[torch.Tensor],
|
| 508 |
+
residual_scale: Optional[torch.Tensor] = None,
|
| 509 |
+
) -> torch.Tensor:
|
| 510 |
+
if not normalized_sources:
|
| 511 |
+
return target_hidden
|
| 512 |
+
update = normalized_sources[-1] * self.beta.to(
|
| 513 |
+
dtype=target_hidden.dtype,
|
| 514 |
+
device=target_hidden.device,
|
| 515 |
+
)
|
| 516 |
+
if residual_scale is not None:
|
| 517 |
+
update = update * residual_scale.to(update.dtype)
|
| 518 |
+
return target_hidden + update.to(target_hidden.dtype)
|
| 519 |
+
|
| 520 |
+
|
| 521 |
+
class NCPFinalConceptRoute(nn.Module):
|
| 522 |
+
def __init__(self, hidden_size: int, eps: float) -> None:
|
| 523 |
+
super().__init__()
|
| 524 |
+
self.concept_norm = nn.LayerNorm(hidden_size, eps=eps)
|
| 525 |
+
self.final_diag = nn.Parameter(torch.empty(hidden_size))
|
| 526 |
+
|
| 527 |
+
def forward(
|
| 528 |
+
self, hidden_states: torch.Tensor, final_concept: torch.Tensor, scale: torch.Tensor
|
| 529 |
+
) -> torch.Tensor:
|
| 530 |
+
update = self.concept_norm(final_concept)
|
| 531 |
+
update = update * self.final_diag.to(update.dtype)
|
| 532 |
+
return hidden_states + update * scale.to(update.dtype)
|
| 533 |
+
|
| 534 |
+
|
| 535 |
+
class NCPFinalConceptCumsumRoute(nn.Module):
|
| 536 |
+
"""LayerNorm plus the scalar final-concept route used in cumsum mode."""
|
| 537 |
+
|
| 538 |
+
def __init__(self, hidden_size: int, eps: float) -> None:
|
| 539 |
+
super().__init__()
|
| 540 |
+
self.concept_norm = nn.LayerNorm(hidden_size, eps=eps)
|
| 541 |
+
self.final_beta = nn.Parameter(torch.empty(()))
|
| 542 |
+
|
| 543 |
+
def forward(
|
| 544 |
+
self, hidden_states: torch.Tensor, final_concept: torch.Tensor, scale: torch.Tensor
|
| 545 |
+
) -> torch.Tensor:
|
| 546 |
+
update = self.concept_norm(final_concept)
|
| 547 |
+
update = update * self.final_beta.to(update.dtype)
|
| 548 |
+
return hidden_states + update * scale.to(update.dtype)
|
| 549 |
+
|
| 550 |
+
|
| 551 |
+
class NCPDDTwoRouteAdd(nn.Module):
|
| 552 |
+
def __init__(self, hidden_size: int, decoder_layers: int, eps: float) -> None:
|
| 553 |
+
super().__init__()
|
| 554 |
+
self.decoder_dds = nn.ModuleList(
|
| 555 |
+
[
|
| 556 |
+
NCPDepthDD(hidden_size, layer_index, use_softmax=True)
|
| 557 |
+
for layer_index in range(int(decoder_layers))
|
| 558 |
+
]
|
| 559 |
+
)
|
| 560 |
+
self.concept_routes = nn.ModuleList(
|
| 561 |
+
[NCPFinalConceptRoute(hidden_size, eps) for _ in range(int(decoder_layers))]
|
| 562 |
+
)
|
| 563 |
+
|
| 564 |
+
|
| 565 |
+
class NCPCumsumTwoRouteAdd(nn.Module):
|
| 566 |
+
def __init__(self, hidden_size: int, decoder_layers: int, eps: float) -> None:
|
| 567 |
+
super().__init__()
|
| 568 |
+
self.decoder_cumsum_dd = NCPSelfCumsumDD()
|
| 569 |
+
self.concept_routes = nn.ModuleList(
|
| 570 |
+
[NCPFinalConceptCumsumRoute(hidden_size, eps) for _ in range(int(decoder_layers))]
|
| 571 |
+
)
|
| 572 |
+
|
| 573 |
+
|
| 574 |
+
class NCPConceptPredictor(nn.Module):
|
| 575 |
+
def __init__(self, config: NCPOlmo3Config) -> None:
|
| 576 |
+
super().__init__()
|
| 577 |
+
hidden_size = int(config.hidden_size)
|
| 578 |
+
hlm_layers = int(config.conceptlm_special_layers)
|
| 579 |
+
encoder_layers = int(config.conceptlm_encoder_layers)
|
| 580 |
+
self.hlm_block = NCPTower(config, hlm_layers, final_layernorm=True)
|
| 581 |
+
self.prediction_heads = nn.ModuleList(
|
| 582 |
+
[
|
| 583 |
+
nn.Linear(hidden_size, int(config.conceptlm_v22_vq_codebook_size))
|
| 584 |
+
for _ in range(int(config.conceptlm_v22_vq_num_codebooks))
|
| 585 |
+
]
|
| 586 |
+
)
|
| 587 |
+
cumsum = str(config.conceptlm_v21_dd_self_dd_mode) == "cumsum"
|
| 588 |
+
self.concept_self_dd = (
|
| 589 |
+
NCPSelfCumsumDD() if cumsum else NCPSelfDD(hidden_size, hlm_layers)
|
| 590 |
+
)
|
| 591 |
+
route_cls = NCPCrossCumsumRoute if cumsum else None
|
| 592 |
+
self.concept_read_encoder_routes = nn.ModuleList(
|
| 593 |
+
[
|
| 594 |
+
route_cls()
|
| 595 |
+
if route_cls is not None
|
| 596 |
+
else NCPDiagResidualRoute(hidden_size, encoder_layers - 1)
|
| 597 |
+
for _ in range(hlm_layers)
|
| 598 |
+
]
|
| 599 |
+
)
|
| 600 |
+
self.concept_read_encoder_shared_source_norm = nn.LayerNorm(
|
| 601 |
+
hidden_size, eps=float(config.layernorm_epsilon)
|
| 602 |
+
)
|
| 603 |
+
|
| 604 |
+
|
| 605 |
+
class NCPProductCodebook(nn.Module):
|
| 606 |
+
def __init__(self, config: NCPOlmo3Config) -> None:
|
| 607 |
+
super().__init__()
|
| 608 |
+
hidden_size = int(config.hidden_size)
|
| 609 |
+
num_codebooks = int(config.conceptlm_v22_vq_num_codebooks)
|
| 610 |
+
codebook_size = int(config.conceptlm_v22_vq_codebook_size)
|
| 611 |
+
code_dim = hidden_size // num_codebooks
|
| 612 |
+
self.codebook = nn.ParameterList(
|
| 613 |
+
[nn.Parameter(torch.empty(codebook_size, code_dim)) for _ in range(num_codebooks)]
|
| 614 |
+
)
|
| 615 |
+
|
| 616 |
+
|
| 617 |
+
class NCPOlmo3ForCausalLM(PreTrainedModel, GenerationMixin):
|
| 618 |
+
config_class = NCPOlmo3Config
|
| 619 |
+
base_model_prefix = ""
|
| 620 |
+
main_input_name = "input_ids"
|
| 621 |
+
_supports_cache_class = False
|
| 622 |
+
_keys_to_ignore_on_load_unexpected = [r".*\._extra_state$"]
|
| 623 |
+
supports_gradient_checkpointing = False
|
| 624 |
+
|
| 625 |
+
def __init__(self, config: NCPOlmo3Config) -> None:
|
| 626 |
+
super().__init__(config)
|
| 627 |
+
hidden_size = int(config.hidden_size)
|
| 628 |
+
encoder_layers = int(config.conceptlm_encoder_layers)
|
| 629 |
+
decoder_layers = int(config.conceptlm_decoder_layers)
|
| 630 |
+
hlm_layers = int(config.conceptlm_special_layers)
|
| 631 |
+
eps = float(config.layernorm_epsilon)
|
| 632 |
+
|
| 633 |
+
self.embed_tokens = nn.Embedding(int(config.vocab_size), hidden_size)
|
| 634 |
+
self.encoder = NCPTower(config, encoder_layers, final_layernorm=False)
|
| 635 |
+
self.decoder = NCPTower(config, decoder_layers, final_layernorm=True)
|
| 636 |
+
self.lm_head = nn.Linear(hidden_size, int(config.vocab_size), bias=False)
|
| 637 |
+
|
| 638 |
+
self.concept_vq_input_norm = nn.LayerNorm(hidden_size, eps=eps)
|
| 639 |
+
self.concept_quantizer = NCPProductCodebook(config)
|
| 640 |
+
self.concept_predictor = NCPConceptPredictor(config)
|
| 641 |
+
|
| 642 |
+
self.cumsum_routes = str(config.conceptlm_v21_dd_self_dd_mode) == "cumsum"
|
| 643 |
+
self.dd_encoder_self_dd = (
|
| 644 |
+
NCPSelfCumsumDD()
|
| 645 |
+
if self.cumsum_routes
|
| 646 |
+
else NCPSelfDD(hidden_size, encoder_layers)
|
| 647 |
+
)
|
| 648 |
+
self.dd_two_route_add = (
|
| 649 |
+
NCPCumsumTwoRouteAdd(hidden_size, decoder_layers, eps)
|
| 650 |
+
if self.cumsum_routes
|
| 651 |
+
else NCPDDTwoRouteAdd(hidden_size, decoder_layers, eps)
|
| 652 |
+
)
|
| 653 |
+
self.decoder_read_encoder_routes = nn.ModuleList(
|
| 654 |
+
[
|
| 655 |
+
NCPCrossCumsumRoute()
|
| 656 |
+
if self.cumsum_routes
|
| 657 |
+
else NCPDiagResidualRoute(hidden_size, encoder_layers)
|
| 658 |
+
for _ in range(decoder_layers)
|
| 659 |
+
]
|
| 660 |
+
)
|
| 661 |
+
self.decoder_read_encoder_shared_source_norm = nn.LayerNorm(hidden_size, eps=eps)
|
| 662 |
+
self.decoder_read_concept_routes = nn.ModuleList(
|
| 663 |
+
[
|
| 664 |
+
NCPCrossCumsumRoute()
|
| 665 |
+
if self.cumsum_routes
|
| 666 |
+
else NCPDiagResidualRoute(hidden_size, hlm_layers)
|
| 667 |
+
for _ in range(decoder_layers)
|
| 668 |
+
]
|
| 669 |
+
)
|
| 670 |
+
self.decoder_read_concept_shared_source_norm = nn.LayerNorm(hidden_size, eps=eps)
|
| 671 |
+
self.fusion_tok_norm = nn.LayerNorm(hidden_size, eps=eps)
|
| 672 |
+
self.fusion_hl_norm = nn.LayerNorm(hidden_size, eps=eps)
|
| 673 |
+
self.fusion_norm_alpha = nn.Parameter(torch.empty(()))
|
| 674 |
+
self.final_read_concept_gate_logits = nn.Parameter(torch.empty(decoder_layers, 2))
|
| 675 |
+
# Required by Transformers 5.x to finalize tied-weight metadata.
|
| 676 |
+
# `from_pretrained` constructs under its no-init context, so this does
|
| 677 |
+
# not rewrite checkpoint values before SafeTensors loading.
|
| 678 |
+
self.post_init()
|
| 679 |
+
|
| 680 |
+
def get_input_embeddings(self) -> nn.Module:
|
| 681 |
+
return self.embed_tokens
|
| 682 |
+
|
| 683 |
+
def set_input_embeddings(self, value: nn.Module) -> None:
|
| 684 |
+
self.embed_tokens = value
|
| 685 |
+
|
| 686 |
+
def get_output_embeddings(self) -> nn.Module:
|
| 687 |
+
return self.lm_head
|
| 688 |
+
|
| 689 |
+
def set_output_embeddings(self, new_embeddings: nn.Module) -> None:
|
| 690 |
+
self.lm_head = new_embeddings
|
| 691 |
+
|
| 692 |
+
def prepare_inputs_for_generation(
|
| 693 |
+
self,
|
| 694 |
+
input_ids: torch.Tensor,
|
| 695 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 696 |
+
past_key_values: Any = None,
|
| 697 |
+
use_cache: Optional[bool] = None,
|
| 698 |
+
**kwargs: Any,
|
| 699 |
+
) -> dict[str, Any]:
|
| 700 |
+
has_past = isinstance(past_key_values, dict)
|
| 701 |
+
if past_key_values is not None and not has_past:
|
| 702 |
+
get_seq_length = getattr(past_key_values, "get_seq_length", None)
|
| 703 |
+
has_past = bool(callable(get_seq_length) and int(get_seq_length()) > 0)
|
| 704 |
+
if has_past:
|
| 705 |
+
input_ids = input_ids[:, -1:]
|
| 706 |
+
return {
|
| 707 |
+
"input_ids": input_ids,
|
| 708 |
+
"attention_mask": attention_mask,
|
| 709 |
+
"past_key_values": past_key_values,
|
| 710 |
+
"use_cache": (bool(self.config.use_cache) if use_cache is None else bool(use_cache)),
|
| 711 |
+
}
|
| 712 |
+
|
| 713 |
+
def _run_tower_layer(
|
| 714 |
+
self,
|
| 715 |
+
layer: NCPOlmo3Layer,
|
| 716 |
+
hidden_states: torch.Tensor,
|
| 717 |
+
position_ids: torch.Tensor,
|
| 718 |
+
attention_mask: Optional[torch.Tensor],
|
| 719 |
+
past_key_value: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
|
| 720 |
+
use_cache: bool = False,
|
| 721 |
+
) -> torch.Tensor | tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
|
| 722 |
+
return layer(
|
| 723 |
+
hidden_states,
|
| 724 |
+
position_ids,
|
| 725 |
+
attention_mask,
|
| 726 |
+
past_key_value=past_key_value,
|
| 727 |
+
use_cache=use_cache,
|
| 728 |
+
)
|
| 729 |
+
|
| 730 |
+
def _apply_self_route(
|
| 731 |
+
self,
|
| 732 |
+
route: NCPSelfDD | NCPSelfCumsumDD,
|
| 733 |
+
layer_index: int,
|
| 734 |
+
raw_output: torch.Tensor,
|
| 735 |
+
history: list[torch.Tensor],
|
| 736 |
+
cumsum_state: Optional[torch.Tensor],
|
| 737 |
+
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
|
| 738 |
+
if isinstance(route, NCPSelfCumsumDD):
|
| 739 |
+
if cumsum_state is None:
|
| 740 |
+
raise RuntimeError("cumsum route state is missing")
|
| 741 |
+
return route(raw_output, cumsum_state)
|
| 742 |
+
return route.depth_dds[layer_index](raw_output, history), cumsum_state
|
| 743 |
+
|
| 744 |
+
def _apply_decoder_self_route(
|
| 745 |
+
self,
|
| 746 |
+
layer_index: int,
|
| 747 |
+
raw_output: torch.Tensor,
|
| 748 |
+
history: list[torch.Tensor],
|
| 749 |
+
cumsum_state: Optional[torch.Tensor],
|
| 750 |
+
) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
|
| 751 |
+
if self.cumsum_routes:
|
| 752 |
+
if cumsum_state is None:
|
| 753 |
+
raise RuntimeError("decoder cumsum route state is missing")
|
| 754 |
+
return self.dd_two_route_add.decoder_cumsum_dd(raw_output, cumsum_state)
|
| 755 |
+
return (
|
| 756 |
+
self.dd_two_route_add.decoder_dds[layer_index](raw_output, history),
|
| 757 |
+
cumsum_state,
|
| 758 |
+
)
|
| 759 |
+
|
| 760 |
+
def _predict_concepts(
|
| 761 |
+
self,
|
| 762 |
+
encoder_hidden: torch.Tensor,
|
| 763 |
+
encoder_raw_layers: list[torch.Tensor],
|
| 764 |
+
*,
|
| 765 |
+
use_cache: bool = False,
|
| 766 |
+
) -> tuple[torch.Tensor, list[torch.Tensor], dict[str, Any]]:
|
| 767 |
+
batch_size, sequence_length, hidden_size = encoder_hidden.shape
|
| 768 |
+
chunk_size = int(self.config.conceptlm_chunk_size)
|
| 769 |
+
num_chunks = sequence_length // chunk_size
|
| 770 |
+
if num_chunks == 0:
|
| 771 |
+
zero = encoder_hidden.new_zeros(batch_size, sequence_length, hidden_size)
|
| 772 |
+
raw_zeros = [zero.clone() for _ in range(int(self.config.conceptlm_special_layers))]
|
| 773 |
+
return (
|
| 774 |
+
zero,
|
| 775 |
+
raw_zeros,
|
| 776 |
+
{
|
| 777 |
+
"hlm": tuple(None for _ in range(int(self.config.conceptlm_special_layers))),
|
| 778 |
+
"latest_concept": zero[:, :1],
|
| 779 |
+
"latest_concept_raw": tuple(item[:, :1] for item in raw_zeros),
|
| 780 |
+
"completed_chunks": 0,
|
| 781 |
+
},
|
| 782 |
+
)
|
| 783 |
+
|
| 784 |
+
completed_tokens = num_chunks * chunk_size
|
| 785 |
+
encoder_chunks = (
|
| 786 |
+
encoder_hidden[:, :completed_tokens]
|
| 787 |
+
.reshape(batch_size, num_chunks, chunk_size, hidden_size)
|
| 788 |
+
.mean(dim=2)
|
| 789 |
+
)
|
| 790 |
+
layer_chunks = [
|
| 791 |
+
states[:, :completed_tokens]
|
| 792 |
+
.reshape(batch_size, num_chunks, chunk_size, hidden_size)
|
| 793 |
+
.mean(dim=2)
|
| 794 |
+
for states in encoder_raw_layers
|
| 795 |
+
]
|
| 796 |
+
|
| 797 |
+
hidden_states = self.concept_vq_input_norm(encoder_chunks)
|
| 798 |
+
encoder_sources = [
|
| 799 |
+
self.concept_predictor.concept_read_encoder_shared_source_norm(source)
|
| 800 |
+
for source in layer_chunks[:-1]
|
| 801 |
+
]
|
| 802 |
+
concept_position_ids = (
|
| 803 |
+
torch.arange(num_chunks, device=encoder_hidden.device, dtype=torch.long)
|
| 804 |
+
.unsqueeze(0)
|
| 805 |
+
.expand(batch_size, -1)
|
| 806 |
+
)
|
| 807 |
+
concept_attention_mask = torch.ones(
|
| 808 |
+
batch_size, num_chunks, device=encoder_hidden.device, dtype=torch.bool
|
| 809 |
+
)
|
| 810 |
+
history = [hidden_states]
|
| 811 |
+
cumsum_state = hidden_states if self.cumsum_routes else None
|
| 812 |
+
raw_layers: list[torch.Tensor] = []
|
| 813 |
+
hlm_present: list[Optional[tuple[torch.Tensor, torch.Tensor]]] = []
|
| 814 |
+
for layer_index, layer in enumerate(self.concept_predictor.hlm_block.layers):
|
| 815 |
+
layer_output = self._run_tower_layer(
|
| 816 |
+
layer,
|
| 817 |
+
hidden_states,
|
| 818 |
+
concept_position_ids,
|
| 819 |
+
concept_attention_mask,
|
| 820 |
+
use_cache=use_cache,
|
| 821 |
+
)
|
| 822 |
+
if use_cache:
|
| 823 |
+
raw_output, present_key_value = layer_output
|
| 824 |
+
hlm_present.append(present_key_value)
|
| 825 |
+
else:
|
| 826 |
+
raw_output = layer_output
|
| 827 |
+
raw_layers.append(raw_output)
|
| 828 |
+
history.append(raw_output)
|
| 829 |
+
hidden_states, cumsum_state = self._apply_self_route(
|
| 830 |
+
self.concept_predictor.concept_self_dd,
|
| 831 |
+
layer_index,
|
| 832 |
+
raw_output,
|
| 833 |
+
history,
|
| 834 |
+
cumsum_state,
|
| 835 |
+
)
|
| 836 |
+
hidden_states = self.concept_predictor.concept_read_encoder_routes[layer_index](
|
| 837 |
+
hidden_states, encoder_sources
|
| 838 |
+
)
|
| 839 |
+
final_norm = self.concept_predictor.hlm_block.norm
|
| 840 |
+
if final_norm is None:
|
| 841 |
+
raise RuntimeError("HLM norm is missing")
|
| 842 |
+
hidden_states = final_norm(hidden_states)
|
| 843 |
+
logits = torch.stack(
|
| 844 |
+
[head(hidden_states) for head in self.concept_predictor.prediction_heads], dim=2
|
| 845 |
+
)
|
| 846 |
+
codebook = torch.stack(tuple(self.concept_quantizer.codebook), dim=0).to(logits.dtype)
|
| 847 |
+
predicted = torch.einsum("bshk,hkd->bshd", logits, codebook).reshape(
|
| 848 |
+
batch_size, num_chunks, hidden_size
|
| 849 |
+
)
|
| 850 |
+
|
| 851 |
+
token_positions = torch.arange(sequence_length, device=encoder_hidden.device)
|
| 852 |
+
concept_indices = (token_positions + 1) // chunk_size - 1
|
| 853 |
+
valid = concept_indices >= 0
|
| 854 |
+
safe_indices = concept_indices.clamp(min=0, max=num_chunks - 1)
|
| 855 |
+
|
| 856 |
+
token_concepts = encoder_hidden.new_zeros(batch_size, sequence_length, hidden_size)
|
| 857 |
+
token_concepts[:, valid] = predicted[:, safe_indices[valid]].to(token_concepts.dtype)
|
| 858 |
+
token_raw_layers = []
|
| 859 |
+
for raw_layer in raw_layers:
|
| 860 |
+
expanded = encoder_hidden.new_zeros(batch_size, sequence_length, hidden_size)
|
| 861 |
+
expanded[:, valid] = raw_layer[:, safe_indices[valid]].to(expanded.dtype)
|
| 862 |
+
token_raw_layers.append(expanded)
|
| 863 |
+
return (
|
| 864 |
+
token_concepts,
|
| 865 |
+
token_raw_layers,
|
| 866 |
+
{
|
| 867 |
+
"hlm": (tuple(hlm_present) if use_cache else tuple(None for _ in raw_layers)),
|
| 868 |
+
"latest_concept": predicted[:, -1:],
|
| 869 |
+
"latest_concept_raw": tuple(raw_layer[:, -1:] for raw_layer in raw_layers),
|
| 870 |
+
"completed_chunks": num_chunks,
|
| 871 |
+
},
|
| 872 |
+
)
|
| 873 |
+
|
| 874 |
+
def _predict_incremental_concept(
|
| 875 |
+
self,
|
| 876 |
+
chunk_encoder_hidden: torch.Tensor,
|
| 877 |
+
chunk_encoder_raw_layers: list[torch.Tensor],
|
| 878 |
+
*,
|
| 879 |
+
completed_chunks: int,
|
| 880 |
+
hlm_past: tuple[Optional[tuple[torch.Tensor, torch.Tensor]], ...],
|
| 881 |
+
) -> tuple[torch.Tensor, list[torch.Tensor], tuple[tuple[torch.Tensor, torch.Tensor], ...]]:
|
| 882 |
+
batch_size = int(chunk_encoder_hidden.shape[0])
|
| 883 |
+
hidden_states = self.concept_vq_input_norm(chunk_encoder_hidden)
|
| 884 |
+
encoder_sources = [
|
| 885 |
+
self.concept_predictor.concept_read_encoder_shared_source_norm(source)
|
| 886 |
+
for source in chunk_encoder_raw_layers[:-1]
|
| 887 |
+
]
|
| 888 |
+
position_ids = torch.full(
|
| 889 |
+
(batch_size, 1), int(completed_chunks), dtype=torch.long, device=hidden_states.device
|
| 890 |
+
)
|
| 891 |
+
attention_mask = torch.ones(
|
| 892 |
+
batch_size, int(completed_chunks) + 1, dtype=torch.bool, device=hidden_states.device
|
| 893 |
+
)
|
| 894 |
+
history = [hidden_states]
|
| 895 |
+
cumsum_state = hidden_states if self.cumsum_routes else None
|
| 896 |
+
raw_layers: list[torch.Tensor] = []
|
| 897 |
+
present_layers: list[tuple[torch.Tensor, torch.Tensor]] = []
|
| 898 |
+
for layer_index, layer in enumerate(self.concept_predictor.hlm_block.layers):
|
| 899 |
+
layer_output = self._run_tower_layer(
|
| 900 |
+
layer,
|
| 901 |
+
hidden_states,
|
| 902 |
+
position_ids,
|
| 903 |
+
attention_mask,
|
| 904 |
+
past_key_value=hlm_past[layer_index],
|
| 905 |
+
use_cache=True,
|
| 906 |
+
)
|
| 907 |
+
raw_output, present_key_value = layer_output
|
| 908 |
+
raw_layers.append(raw_output)
|
| 909 |
+
present_layers.append(present_key_value)
|
| 910 |
+
history.append(raw_output)
|
| 911 |
+
hidden_states, cumsum_state = self._apply_self_route(
|
| 912 |
+
self.concept_predictor.concept_self_dd,
|
| 913 |
+
layer_index,
|
| 914 |
+
raw_output,
|
| 915 |
+
history,
|
| 916 |
+
cumsum_state,
|
| 917 |
+
)
|
| 918 |
+
hidden_states = self.concept_predictor.concept_read_encoder_routes[layer_index](
|
| 919 |
+
hidden_states, encoder_sources
|
| 920 |
+
)
|
| 921 |
+
final_norm = self.concept_predictor.hlm_block.norm
|
| 922 |
+
if final_norm is None:
|
| 923 |
+
raise RuntimeError("HLM norm is missing")
|
| 924 |
+
hidden_states = final_norm(hidden_states)
|
| 925 |
+
logits = torch.stack(
|
| 926 |
+
[head(hidden_states) for head in self.concept_predictor.prediction_heads], dim=2
|
| 927 |
+
)
|
| 928 |
+
codebook = torch.stack(tuple(self.concept_quantizer.codebook), dim=0).to(logits.dtype)
|
| 929 |
+
predicted = torch.einsum("bshk,hkd->bshd", logits, codebook).reshape(batch_size, 1, -1)
|
| 930 |
+
return predicted, raw_layers, tuple(present_layers)
|
| 931 |
+
|
| 932 |
+
def _forward_incremental(
|
| 933 |
+
self,
|
| 934 |
+
inputs_embeds: torch.Tensor,
|
| 935 |
+
attention_mask: torch.Tensor,
|
| 936 |
+
position_ids: torch.Tensor,
|
| 937 |
+
cache: dict[str, Any],
|
| 938 |
+
) -> tuple[torch.Tensor, dict[str, Any]]:
|
| 939 |
+
if int(inputs_embeds.shape[1]) != 1:
|
| 940 |
+
raise ValueError("cached inference expects exactly one new token")
|
| 941 |
+
past_length = int(cache["sequence_length"])
|
| 942 |
+
new_length = past_length + 1
|
| 943 |
+
hidden_states = inputs_embeds
|
| 944 |
+
|
| 945 |
+
encoder_history = [hidden_states]
|
| 946 |
+
encoder_cumsum_state = hidden_states if self.cumsum_routes else None
|
| 947 |
+
encoder_raw_layers: list[torch.Tensor] = []
|
| 948 |
+
encoder_present: list[tuple[torch.Tensor, torch.Tensor]] = []
|
| 949 |
+
for layer_index, layer in enumerate(self.encoder.layers):
|
| 950 |
+
layer_output = self._run_tower_layer(
|
| 951 |
+
layer,
|
| 952 |
+
hidden_states,
|
| 953 |
+
position_ids,
|
| 954 |
+
attention_mask,
|
| 955 |
+
past_key_value=cache["encoder"][layer_index],
|
| 956 |
+
use_cache=True,
|
| 957 |
+
)
|
| 958 |
+
raw_output, present_key_value = layer_output
|
| 959 |
+
encoder_raw_layers.append(raw_output)
|
| 960 |
+
encoder_present.append(present_key_value)
|
| 961 |
+
encoder_history.append(raw_output)
|
| 962 |
+
hidden_states, encoder_cumsum_state = self._apply_self_route(
|
| 963 |
+
self.dd_encoder_self_dd,
|
| 964 |
+
layer_index,
|
| 965 |
+
raw_output,
|
| 966 |
+
encoder_history,
|
| 967 |
+
encoder_cumsum_state,
|
| 968 |
+
)
|
| 969 |
+
encoder_hidden = hidden_states
|
| 970 |
+
|
| 971 |
+
partial_encoder_sum = cache["partial_encoder_sum"] + encoder_hidden
|
| 972 |
+
partial_encoder_raw_sums = [
|
| 973 |
+
previous + current
|
| 974 |
+
for previous, current in zip(
|
| 975 |
+
cache["partial_encoder_raw_sums"], encoder_raw_layers, strict=True
|
| 976 |
+
)
|
| 977 |
+
]
|
| 978 |
+
chunk_size = int(self.config.conceptlm_chunk_size)
|
| 979 |
+
completed_chunks = int(cache["completed_chunks"])
|
| 980 |
+
if new_length % chunk_size == 0:
|
| 981 |
+
chunk_encoder_hidden = partial_encoder_sum / float(chunk_size)
|
| 982 |
+
chunk_encoder_raw = [value / float(chunk_size) for value in partial_encoder_raw_sums]
|
| 983 |
+
(latest_concept, latest_concept_raw, hlm_present) = self._predict_incremental_concept(
|
| 984 |
+
chunk_encoder_hidden,
|
| 985 |
+
chunk_encoder_raw,
|
| 986 |
+
completed_chunks=completed_chunks,
|
| 987 |
+
hlm_past=cache["hlm"],
|
| 988 |
+
)
|
| 989 |
+
completed_chunks += 1
|
| 990 |
+
partial_encoder_sum = torch.zeros_like(partial_encoder_sum)
|
| 991 |
+
partial_encoder_raw_sums = [
|
| 992 |
+
torch.zeros_like(value) for value in partial_encoder_raw_sums
|
| 993 |
+
]
|
| 994 |
+
else:
|
| 995 |
+
latest_concept = cache["latest_concept"]
|
| 996 |
+
latest_concept_raw = list(cache["latest_concept_raw"])
|
| 997 |
+
hlm_present = cache["hlm"]
|
| 998 |
+
|
| 999 |
+
if new_length < chunk_size:
|
| 1000 |
+
final_concepts = torch.zeros_like(encoder_hidden)
|
| 1001 |
+
concept_raw_layers = [
|
| 1002 |
+
torch.zeros_like(encoder_hidden)
|
| 1003 |
+
for _ in range(int(self.config.conceptlm_special_layers))
|
| 1004 |
+
]
|
| 1005 |
+
else:
|
| 1006 |
+
final_concepts = latest_concept.to(encoder_hidden.dtype)
|
| 1007 |
+
concept_raw_layers = [value.to(encoder_hidden.dtype) for value in latest_concept_raw]
|
| 1008 |
+
|
| 1009 |
+
hidden_states = self.fusion_tok_norm(encoder_hidden) + self.fusion_norm_alpha.to(
|
| 1010 |
+
encoder_hidden.dtype
|
| 1011 |
+
) * self.fusion_hl_norm(final_concepts)
|
| 1012 |
+
encoder_sources = [
|
| 1013 |
+
self.decoder_read_encoder_shared_source_norm(source) for source in encoder_raw_layers
|
| 1014 |
+
]
|
| 1015 |
+
concept_sources = [
|
| 1016 |
+
self.decoder_read_concept_shared_source_norm(source) for source in concept_raw_layers
|
| 1017 |
+
]
|
| 1018 |
+
decoder_history = [hidden_states]
|
| 1019 |
+
decoder_cumsum_state = hidden_states if self.cumsum_routes else None
|
| 1020 |
+
decoder_present: list[tuple[torch.Tensor, torch.Tensor]] = []
|
| 1021 |
+
for layer_index, layer in enumerate(self.decoder.layers):
|
| 1022 |
+
layer_output = self._run_tower_layer(
|
| 1023 |
+
layer,
|
| 1024 |
+
hidden_states,
|
| 1025 |
+
position_ids,
|
| 1026 |
+
attention_mask,
|
| 1027 |
+
past_key_value=cache["decoder"][layer_index],
|
| 1028 |
+
use_cache=True,
|
| 1029 |
+
)
|
| 1030 |
+
raw_output, present_key_value = layer_output
|
| 1031 |
+
decoder_present.append(present_key_value)
|
| 1032 |
+
decoder_history.append(raw_output)
|
| 1033 |
+
gate = self.final_read_concept_gate_logits[layer_index].float().softmax(dim=-1)
|
| 1034 |
+
hidden_states, decoder_cumsum_state = self._apply_decoder_self_route(
|
| 1035 |
+
layer_index,
|
| 1036 |
+
raw_output,
|
| 1037 |
+
decoder_history,
|
| 1038 |
+
decoder_cumsum_state,
|
| 1039 |
+
)
|
| 1040 |
+
hidden_states = self.dd_two_route_add.concept_routes[layer_index](
|
| 1041 |
+
hidden_states, final_concepts, gate[0]
|
| 1042 |
+
)
|
| 1043 |
+
hidden_states = self.decoder_read_encoder_routes[layer_index](
|
| 1044 |
+
hidden_states, encoder_sources
|
| 1045 |
+
)
|
| 1046 |
+
hidden_states = self.decoder_read_concept_routes[layer_index](
|
| 1047 |
+
hidden_states, concept_sources, residual_scale=gate[1]
|
| 1048 |
+
)
|
| 1049 |
+
final_norm = self.decoder.norm
|
| 1050 |
+
if final_norm is None:
|
| 1051 |
+
raise RuntimeError("decoder norm is missing")
|
| 1052 |
+
logits = self.lm_head(final_norm(hidden_states))
|
| 1053 |
+
updated_cache = {
|
| 1054 |
+
"sequence_length": new_length,
|
| 1055 |
+
"encoder": tuple(encoder_present),
|
| 1056 |
+
"hlm": tuple(hlm_present),
|
| 1057 |
+
"decoder": tuple(decoder_present),
|
| 1058 |
+
"partial_encoder_sum": partial_encoder_sum,
|
| 1059 |
+
"partial_encoder_raw_sums": tuple(partial_encoder_raw_sums),
|
| 1060 |
+
"latest_concept": latest_concept,
|
| 1061 |
+
"latest_concept_raw": tuple(latest_concept_raw),
|
| 1062 |
+
"completed_chunks": completed_chunks,
|
| 1063 |
+
}
|
| 1064 |
+
return logits, updated_cache
|
| 1065 |
+
|
| 1066 |
+
def forward(
|
| 1067 |
+
self,
|
| 1068 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 1069 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1070 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 1071 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 1072 |
+
labels: Optional[torch.Tensor] = None,
|
| 1073 |
+
use_cache: Optional[bool] = None,
|
| 1074 |
+
past_key_values: Any = None,
|
| 1075 |
+
return_dict: Optional[bool] = None,
|
| 1076 |
+
**kwargs: Any,
|
| 1077 |
+
) -> CausalLMOutputWithPast | tuple[torch.Tensor, ...]:
|
| 1078 |
+
del kwargs
|
| 1079 |
+
use_cache = bool(self.config.use_cache) if use_cache is None else bool(use_cache)
|
| 1080 |
+
if inputs_embeds is None:
|
| 1081 |
+
if input_ids is None:
|
| 1082 |
+
raise ValueError("input_ids or inputs_embeds is required")
|
| 1083 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
| 1084 |
+
if inputs_embeds.ndim != 3 or inputs_embeds.shape[0] != 1:
|
| 1085 |
+
raise ValueError("the initial standalone correctness path requires batch_size=1")
|
| 1086 |
+
batch_size, sequence_length, _ = inputs_embeds.shape
|
| 1087 |
+
if attention_mask is None:
|
| 1088 |
+
attention_mask = torch.ones(
|
| 1089 |
+
batch_size, sequence_length, dtype=torch.bool, device=inputs_embeds.device
|
| 1090 |
+
)
|
| 1091 |
+
if position_ids is None:
|
| 1092 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
| 1093 |
+
position_ids.masked_fill_(~attention_mask.bool(), 0)
|
| 1094 |
+
position_ids = position_ids[:, -sequence_length:]
|
| 1095 |
+
|
| 1096 |
+
if past_key_values is not None and not isinstance(past_key_values, dict):
|
| 1097 |
+
get_seq_length = getattr(past_key_values, "get_seq_length", None)
|
| 1098 |
+
if callable(get_seq_length) and int(get_seq_length()) == 0:
|
| 1099 |
+
# GenerationMixin may provision an empty DynamicCache before
|
| 1100 |
+
# the first model call. The NCP graph replaces it with its
|
| 1101 |
+
# richer cache dictionary in the returned output.
|
| 1102 |
+
past_key_values = None
|
| 1103 |
+
else:
|
| 1104 |
+
raise TypeError("NCP standalone cache must be a dictionary")
|
| 1105 |
+
|
| 1106 |
+
if past_key_values is not None:
|
| 1107 |
+
logits, updated_cache = self._forward_incremental(
|
| 1108 |
+
inputs_embeds, attention_mask, position_ids, past_key_values
|
| 1109 |
+
)
|
| 1110 |
+
output = CausalLMOutputWithPast(loss=None, logits=logits, past_key_values=updated_cache)
|
| 1111 |
+
if return_dict is False:
|
| 1112 |
+
return (logits, updated_cache)
|
| 1113 |
+
return output
|
| 1114 |
+
|
| 1115 |
+
hidden_states = inputs_embeds
|
| 1116 |
+
encoder_history = [hidden_states]
|
| 1117 |
+
encoder_cumsum_state = hidden_states if self.cumsum_routes else None
|
| 1118 |
+
encoder_raw_layers: list[torch.Tensor] = []
|
| 1119 |
+
encoder_present: list[tuple[torch.Tensor, torch.Tensor]] = []
|
| 1120 |
+
for layer_index, layer in enumerate(self.encoder.layers):
|
| 1121 |
+
layer_output = self._run_tower_layer(
|
| 1122 |
+
layer, hidden_states, position_ids, attention_mask, use_cache=use_cache
|
| 1123 |
+
)
|
| 1124 |
+
if use_cache:
|
| 1125 |
+
raw_output, present_key_value = layer_output
|
| 1126 |
+
encoder_present.append(present_key_value)
|
| 1127 |
+
else:
|
| 1128 |
+
raw_output = layer_output
|
| 1129 |
+
encoder_raw_layers.append(raw_output)
|
| 1130 |
+
encoder_history.append(raw_output)
|
| 1131 |
+
hidden_states, encoder_cumsum_state = self._apply_self_route(
|
| 1132 |
+
self.dd_encoder_self_dd,
|
| 1133 |
+
layer_index,
|
| 1134 |
+
raw_output,
|
| 1135 |
+
encoder_history,
|
| 1136 |
+
encoder_cumsum_state,
|
| 1137 |
+
)
|
| 1138 |
+
encoder_hidden = hidden_states
|
| 1139 |
+
|
| 1140 |
+
(final_concepts, concept_raw_layers, concept_state) = self._predict_concepts(
|
| 1141 |
+
encoder_hidden, encoder_raw_layers, use_cache=use_cache
|
| 1142 |
+
)
|
| 1143 |
+
hidden_states = self.fusion_tok_norm(encoder_hidden) + self.fusion_norm_alpha.to(
|
| 1144 |
+
encoder_hidden.dtype
|
| 1145 |
+
) * self.fusion_hl_norm(final_concepts)
|
| 1146 |
+
encoder_sources = [
|
| 1147 |
+
self.decoder_read_encoder_shared_source_norm(source) for source in encoder_raw_layers
|
| 1148 |
+
]
|
| 1149 |
+
concept_sources = [
|
| 1150 |
+
self.decoder_read_concept_shared_source_norm(source) for source in concept_raw_layers
|
| 1151 |
+
]
|
| 1152 |
+
|
| 1153 |
+
decoder_history = [hidden_states]
|
| 1154 |
+
decoder_cumsum_state = hidden_states if self.cumsum_routes else None
|
| 1155 |
+
decoder_present: list[tuple[torch.Tensor, torch.Tensor]] = []
|
| 1156 |
+
for layer_index, layer in enumerate(self.decoder.layers):
|
| 1157 |
+
layer_output = self._run_tower_layer(
|
| 1158 |
+
layer, hidden_states, position_ids, attention_mask, use_cache=use_cache
|
| 1159 |
+
)
|
| 1160 |
+
if use_cache:
|
| 1161 |
+
raw_output, present_key_value = layer_output
|
| 1162 |
+
decoder_present.append(present_key_value)
|
| 1163 |
+
else:
|
| 1164 |
+
raw_output = layer_output
|
| 1165 |
+
decoder_history.append(raw_output)
|
| 1166 |
+
gate = self.final_read_concept_gate_logits[layer_index].float().softmax(dim=-1)
|
| 1167 |
+
hidden_states, decoder_cumsum_state = self._apply_decoder_self_route(
|
| 1168 |
+
layer_index,
|
| 1169 |
+
raw_output,
|
| 1170 |
+
decoder_history,
|
| 1171 |
+
decoder_cumsum_state,
|
| 1172 |
+
)
|
| 1173 |
+
hidden_states = self.dd_two_route_add.concept_routes[layer_index](
|
| 1174 |
+
hidden_states, final_concepts, gate[0]
|
| 1175 |
+
)
|
| 1176 |
+
hidden_states = self.decoder_read_encoder_routes[layer_index](
|
| 1177 |
+
hidden_states, encoder_sources
|
| 1178 |
+
)
|
| 1179 |
+
hidden_states = self.decoder_read_concept_routes[layer_index](
|
| 1180 |
+
hidden_states, concept_sources, residual_scale=gate[1]
|
| 1181 |
+
)
|
| 1182 |
+
final_norm = self.decoder.norm
|
| 1183 |
+
if final_norm is None:
|
| 1184 |
+
raise RuntimeError("decoder norm is missing")
|
| 1185 |
+
hidden_states = final_norm(hidden_states)
|
| 1186 |
+
logits = self.lm_head(hidden_states)
|
| 1187 |
+
|
| 1188 |
+
loss = None
|
| 1189 |
+
if labels is not None:
|
| 1190 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
| 1191 |
+
shift_labels = labels[..., 1:].contiguous()
|
| 1192 |
+
loss = F.cross_entropy(
|
| 1193 |
+
shift_logits.float().view(-1, shift_logits.shape[-1]),
|
| 1194 |
+
shift_labels.view(-1),
|
| 1195 |
+
ignore_index=-100,
|
| 1196 |
+
)
|
| 1197 |
+
|
| 1198 |
+
updated_cache = None
|
| 1199 |
+
if use_cache:
|
| 1200 |
+
chunk_size = int(self.config.conceptlm_chunk_size)
|
| 1201 |
+
remainder = sequence_length % chunk_size
|
| 1202 |
+
if remainder:
|
| 1203 |
+
partial_encoder_sum = encoder_hidden[:, -remainder:].sum(dim=1, keepdim=True)
|
| 1204 |
+
partial_encoder_raw_sums = tuple(
|
| 1205 |
+
value[:, -remainder:].sum(dim=1, keepdim=True) for value in encoder_raw_layers
|
| 1206 |
+
)
|
| 1207 |
+
else:
|
| 1208 |
+
partial_encoder_sum = encoder_hidden.new_zeros(
|
| 1209 |
+
batch_size, 1, encoder_hidden.shape[-1]
|
| 1210 |
+
)
|
| 1211 |
+
partial_encoder_raw_sums = tuple(
|
| 1212 |
+
torch.zeros_like(partial_encoder_sum) for _ in encoder_raw_layers
|
| 1213 |
+
)
|
| 1214 |
+
updated_cache = {
|
| 1215 |
+
"sequence_length": sequence_length,
|
| 1216 |
+
"encoder": tuple(encoder_present),
|
| 1217 |
+
"hlm": concept_state["hlm"],
|
| 1218 |
+
"decoder": tuple(decoder_present),
|
| 1219 |
+
"partial_encoder_sum": partial_encoder_sum,
|
| 1220 |
+
"partial_encoder_raw_sums": partial_encoder_raw_sums,
|
| 1221 |
+
"latest_concept": concept_state["latest_concept"],
|
| 1222 |
+
"latest_concept_raw": concept_state["latest_concept_raw"],
|
| 1223 |
+
"completed_chunks": concept_state["completed_chunks"],
|
| 1224 |
+
}
|
| 1225 |
+
output = CausalLMOutputWithPast(loss=loss, logits=logits, past_key_values=updated_cache)
|
| 1226 |
+
if return_dict is False:
|
| 1227 |
+
values: tuple[Any, ...] = (logits, updated_cache) if use_cache else (logits,)
|
| 1228 |
+
return ((loss,) + values) if loss is not None else values
|
| 1229 |
+
return output
|
| 1230 |
+
|
| 1231 |
+
|
| 1232 |
+
ConceptLMV22VQForCausalLM = NCPOlmo3ForCausalLM
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": "<|endoftext|>",
|
| 3 |
+
"eos_token": "<|endoftext|>",
|
| 4 |
+
"pad_token": "<|pad|>",
|
| 5 |
+
"unk_token": "<|endoftext|>"
|
| 6 |
+
}
|
standalone_manifest.json
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"status": "NCP_OLMO3_PURE_HF_READY",
|
| 3 |
+
"architecture": "NCPOlmo3ForCausalLM",
|
| 4 |
+
"model_type": "ncp_olmo3",
|
| 5 |
+
"weight_key_format": "huggingface_state_dict",
|
| 6 |
+
"auto_map": {
|
| 7 |
+
"AutoConfig": "configuration_ncp_olmo3.NCPOlmo3Config",
|
| 8 |
+
"AutoModelForCausalLM": "modeling_ncp_olmo3.NCPOlmo3ForCausalLM"
|
| 9 |
+
},
|
| 10 |
+
"tensor_count": 842,
|
| 11 |
+
"source_checkpoint_root": "/mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/checkpoints",
|
| 12 |
+
"source_checkpoint_step": 1382310,
|
| 13 |
+
"source_checkpoint_mutated": false,
|
| 14 |
+
"hf_backend": "AutoModelForCausalLM.from_pretrained",
|
| 15 |
+
"vllm_backend": "conceptlm_vllm_backend",
|
| 16 |
+
"external_runtime_required": false,
|
| 17 |
+
"position_embedding_type": "rope",
|
| 18 |
+
"max_position_embeddings": 8192,
|
| 19 |
+
"window_size": [
|
| 20 |
+
4096,
|
| 21 |
+
0
|
| 22 |
+
],
|
| 23 |
+
"window_attn_skip_freq": 4,
|
| 24 |
+
"yarn_full_attn_layers_only": null
|
| 25 |
+
}
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,189 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_prefix_space": false,
|
| 3 |
+
"added_tokens_decoder": {
|
| 4 |
+
"100256": {
|
| 5 |
+
"content": "<|extra_id_0|>",
|
| 6 |
+
"lstrip": false,
|
| 7 |
+
"normalized": false,
|
| 8 |
+
"rstrip": false,
|
| 9 |
+
"single_word": false,
|
| 10 |
+
"special": false
|
| 11 |
+
},
|
| 12 |
+
"100257": {
|
| 13 |
+
"content": "<|endoftext|>",
|
| 14 |
+
"lstrip": false,
|
| 15 |
+
"normalized": false,
|
| 16 |
+
"rstrip": false,
|
| 17 |
+
"single_word": false,
|
| 18 |
+
"special": true
|
| 19 |
+
},
|
| 20 |
+
"100258": {
|
| 21 |
+
"content": "<|fim_prefix|>",
|
| 22 |
+
"lstrip": false,
|
| 23 |
+
"normalized": false,
|
| 24 |
+
"rstrip": false,
|
| 25 |
+
"single_word": false,
|
| 26 |
+
"special": true
|
| 27 |
+
},
|
| 28 |
+
"100259": {
|
| 29 |
+
"content": "<|fim_middle|>",
|
| 30 |
+
"lstrip": false,
|
| 31 |
+
"normalized": false,
|
| 32 |
+
"rstrip": false,
|
| 33 |
+
"single_word": false,
|
| 34 |
+
"special": true
|
| 35 |
+
},
|
| 36 |
+
"100260": {
|
| 37 |
+
"content": "<|fim_suffix|>",
|
| 38 |
+
"lstrip": false,
|
| 39 |
+
"normalized": false,
|
| 40 |
+
"rstrip": false,
|
| 41 |
+
"single_word": false,
|
| 42 |
+
"special": true
|
| 43 |
+
},
|
| 44 |
+
"100261": {
|
| 45 |
+
"content": "|||PHONE_NUMBER|||",
|
| 46 |
+
"lstrip": false,
|
| 47 |
+
"normalized": false,
|
| 48 |
+
"rstrip": false,
|
| 49 |
+
"single_word": false,
|
| 50 |
+
"special": false
|
| 51 |
+
},
|
| 52 |
+
"100262": {
|
| 53 |
+
"content": "|||EMAIL_ADDRESS|||",
|
| 54 |
+
"lstrip": false,
|
| 55 |
+
"normalized": false,
|
| 56 |
+
"rstrip": false,
|
| 57 |
+
"single_word": false,
|
| 58 |
+
"special": false
|
| 59 |
+
},
|
| 60 |
+
"100263": {
|
| 61 |
+
"content": "|||IP_ADDRESS|||",
|
| 62 |
+
"lstrip": false,
|
| 63 |
+
"normalized": false,
|
| 64 |
+
"rstrip": false,
|
| 65 |
+
"single_word": false,
|
| 66 |
+
"special": false
|
| 67 |
+
},
|
| 68 |
+
"100264": {
|
| 69 |
+
"content": "<|im_start|>",
|
| 70 |
+
"lstrip": false,
|
| 71 |
+
"normalized": false,
|
| 72 |
+
"rstrip": false,
|
| 73 |
+
"single_word": false,
|
| 74 |
+
"special": true
|
| 75 |
+
},
|
| 76 |
+
"100265": {
|
| 77 |
+
"content": "<|im_end|>",
|
| 78 |
+
"lstrip": false,
|
| 79 |
+
"normalized": false,
|
| 80 |
+
"rstrip": false,
|
| 81 |
+
"single_word": false,
|
| 82 |
+
"special": true
|
| 83 |
+
},
|
| 84 |
+
"100266": {
|
| 85 |
+
"content": "<|extra_id_1|>",
|
| 86 |
+
"lstrip": false,
|
| 87 |
+
"normalized": false,
|
| 88 |
+
"rstrip": false,
|
| 89 |
+
"single_word": false,
|
| 90 |
+
"special": false
|
| 91 |
+
},
|
| 92 |
+
"100267": {
|
| 93 |
+
"content": "<|extra_id_2|>",
|
| 94 |
+
"lstrip": false,
|
| 95 |
+
"normalized": false,
|
| 96 |
+
"rstrip": false,
|
| 97 |
+
"single_word": false,
|
| 98 |
+
"special": false
|
| 99 |
+
},
|
| 100 |
+
"100268": {
|
| 101 |
+
"content": "<|extra_id_3|>",
|
| 102 |
+
"lstrip": false,
|
| 103 |
+
"normalized": false,
|
| 104 |
+
"rstrip": false,
|
| 105 |
+
"single_word": false,
|
| 106 |
+
"special": false
|
| 107 |
+
},
|
| 108 |
+
"100269": {
|
| 109 |
+
"content": "<|extra_id_4|>",
|
| 110 |
+
"lstrip": false,
|
| 111 |
+
"normalized": false,
|
| 112 |
+
"rstrip": false,
|
| 113 |
+
"single_word": false,
|
| 114 |
+
"special": false
|
| 115 |
+
},
|
| 116 |
+
"100270": {
|
| 117 |
+
"content": "<|extra_id_5|>",
|
| 118 |
+
"lstrip": false,
|
| 119 |
+
"normalized": false,
|
| 120 |
+
"rstrip": false,
|
| 121 |
+
"single_word": false,
|
| 122 |
+
"special": false
|
| 123 |
+
},
|
| 124 |
+
"100271": {
|
| 125 |
+
"content": "<|extra_id_6|>",
|
| 126 |
+
"lstrip": false,
|
| 127 |
+
"normalized": false,
|
| 128 |
+
"rstrip": false,
|
| 129 |
+
"single_word": false,
|
| 130 |
+
"special": false
|
| 131 |
+
},
|
| 132 |
+
"100272": {
|
| 133 |
+
"content": "<|extra_id_7|>",
|
| 134 |
+
"lstrip": false,
|
| 135 |
+
"normalized": false,
|
| 136 |
+
"rstrip": false,
|
| 137 |
+
"single_word": false,
|
| 138 |
+
"special": false
|
| 139 |
+
},
|
| 140 |
+
"100273": {
|
| 141 |
+
"content": "<|extra_id_8|>",
|
| 142 |
+
"lstrip": false,
|
| 143 |
+
"normalized": false,
|
| 144 |
+
"rstrip": false,
|
| 145 |
+
"single_word": false,
|
| 146 |
+
"special": false
|
| 147 |
+
},
|
| 148 |
+
"100274": {
|
| 149 |
+
"content": "<|extra_id_9|>",
|
| 150 |
+
"lstrip": false,
|
| 151 |
+
"normalized": false,
|
| 152 |
+
"rstrip": false,
|
| 153 |
+
"single_word": false,
|
| 154 |
+
"special": false
|
| 155 |
+
},
|
| 156 |
+
"100275": {
|
| 157 |
+
"content": "<|extra_id_10|>",
|
| 158 |
+
"lstrip": false,
|
| 159 |
+
"normalized": false,
|
| 160 |
+
"rstrip": false,
|
| 161 |
+
"single_word": false,
|
| 162 |
+
"special": false
|
| 163 |
+
},
|
| 164 |
+
"100276": {
|
| 165 |
+
"content": "<|endofprompt|>",
|
| 166 |
+
"lstrip": false,
|
| 167 |
+
"normalized": false,
|
| 168 |
+
"rstrip": false,
|
| 169 |
+
"single_word": false,
|
| 170 |
+
"special": true
|
| 171 |
+
},
|
| 172 |
+
"100277": {
|
| 173 |
+
"content": "<|pad|>",
|
| 174 |
+
"lstrip": false,
|
| 175 |
+
"normalized": false,
|
| 176 |
+
"rstrip": false,
|
| 177 |
+
"single_word": false,
|
| 178 |
+
"special": true
|
| 179 |
+
}
|
| 180 |
+
},
|
| 181 |
+
"bos_token": "<|endoftext|>",
|
| 182 |
+
"clean_up_tokenization_spaces": false,
|
| 183 |
+
"eos_token": "<|endoftext|>",
|
| 184 |
+
"extra_special_tokens": {},
|
| 185 |
+
"model_max_length": 1000000000000000019884624838656,
|
| 186 |
+
"pad_token": "<|pad|>",
|
| 187 |
+
"tokenizer_class": "TokenizersBackend",
|
| 188 |
+
"unk_token": "<|endoftext|>"
|
| 189 |
+
}
|
training_config.yaml
ADDED
|
@@ -0,0 +1,2805 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
wandb_version: 1
|
| 2 |
+
|
| 3 |
+
_wandb:
|
| 4 |
+
desc: null
|
| 5 |
+
value:
|
| 6 |
+
python_version: 3.12.13
|
| 7 |
+
cli_version: 0.26.1
|
| 8 |
+
framework: huggingface
|
| 9 |
+
huggingface_version: 5.8.0
|
| 10 |
+
is_jupyter_run: false
|
| 11 |
+
is_kaggle_kernel: false
|
| 12 |
+
start_time: 1781149227
|
| 13 |
+
t:
|
| 14 |
+
1:
|
| 15 |
+
- 1
|
| 16 |
+
- 11
|
| 17 |
+
- 49
|
| 18 |
+
2:
|
| 19 |
+
- 1
|
| 20 |
+
- 11
|
| 21 |
+
- 49
|
| 22 |
+
3:
|
| 23 |
+
- 4
|
| 24 |
+
- 13
|
| 25 |
+
- 16
|
| 26 |
+
- 42
|
| 27 |
+
- 61
|
| 28 |
+
4: 3.12.13
|
| 29 |
+
5: 0.26.1
|
| 30 |
+
6: 5.8.0
|
| 31 |
+
13: linux-x86_64
|
| 32 |
+
e:
|
| 33 |
+
lwpdwcz10y13ud608inuzplox070dihm:
|
| 34 |
+
os: Linux-5.15.0-89-generic-x86_64-with-glibc2.35
|
| 35 |
+
python: CPython 3.12.13
|
| 36 |
+
started_at: '2026-06-11T03:40:27.005348Z'
|
| 37 |
+
args:
|
| 38 |
+
- --use-mcore-models
|
| 39 |
+
- --transformer-impl
|
| 40 |
+
- transformer_engine
|
| 41 |
+
- --spec
|
| 42 |
+
- megatron.core.models.gpt.olmo3_layer_specs
|
| 43 |
+
- olmo3_layer_spec
|
| 44 |
+
- --num-layers
|
| 45 |
+
- '32'
|
| 46 |
+
- --hidden-size
|
| 47 |
+
- '4096'
|
| 48 |
+
- --ffn-hidden-size
|
| 49 |
+
- '11008'
|
| 50 |
+
- --num-attention-heads
|
| 51 |
+
- '32'
|
| 52 |
+
- --kv-channels
|
| 53 |
+
- '128'
|
| 54 |
+
- --seq-length
|
| 55 |
+
- '8192'
|
| 56 |
+
- --max-position-embeddings
|
| 57 |
+
- '8192'
|
| 58 |
+
- --position-embedding-type
|
| 59 |
+
- rope
|
| 60 |
+
- --rotary-base
|
| 61 |
+
- '500000'
|
| 62 |
+
- --rotary-percent
|
| 63 |
+
- '1.0'
|
| 64 |
+
- --make-vocab-size-divisible-by
|
| 65 |
+
- '1'
|
| 66 |
+
- --window-size
|
| 67 |
+
- 4096,0
|
| 68 |
+
- --window-attn-skip-freq
|
| 69 |
+
- '4'
|
| 70 |
+
- --qk-layernorm
|
| 71 |
+
- --swiglu
|
| 72 |
+
- --normalization
|
| 73 |
+
- RMSNorm
|
| 74 |
+
- --norm-epsilon
|
| 75 |
+
- 1e-6
|
| 76 |
+
- --attention-dropout
|
| 77 |
+
- '0.0'
|
| 78 |
+
- --hidden-dropout
|
| 79 |
+
- '0.0'
|
| 80 |
+
- --disable-bias-linear
|
| 81 |
+
- --untie-embeddings-and-output-weights
|
| 82 |
+
- --init-method-std
|
| 83 |
+
- '0.02'
|
| 84 |
+
- --init-method-variant
|
| 85 |
+
- scaled_truncated_normal
|
| 86 |
+
- --micro-batch-size
|
| 87 |
+
- '1'
|
| 88 |
+
- --global-batch-size
|
| 89 |
+
- '512'
|
| 90 |
+
- --train-iters
|
| 91 |
+
- '1430512'
|
| 92 |
+
- --lr-decay-iters
|
| 93 |
+
- '1430512'
|
| 94 |
+
- --lr-warmup-iters
|
| 95 |
+
- '2000'
|
| 96 |
+
- --lr
|
| 97 |
+
- '6.0e-5'
|
| 98 |
+
- --min-lr
|
| 99 |
+
- '6.0e-6'
|
| 100 |
+
- --lr-decay-style
|
| 101 |
+
- cosine
|
| 102 |
+
- --clip-grad
|
| 103 |
+
- '1.0'
|
| 104 |
+
- --weight-decay
|
| 105 |
+
- '0.1'
|
| 106 |
+
- --optimizer
|
| 107 |
+
- muon
|
| 108 |
+
- --adam-beta1
|
| 109 |
+
- '0.9'
|
| 110 |
+
- --adam-beta2
|
| 111 |
+
- '0.95'
|
| 112 |
+
- --adam-eps
|
| 113 |
+
- '1.0e-8'
|
| 114 |
+
- --bf16
|
| 115 |
+
- --cross-entropy-loss-fusion
|
| 116 |
+
- --tensor-model-parallel-size
|
| 117 |
+
- '1'
|
| 118 |
+
- --pipeline-model-parallel-size
|
| 119 |
+
- '1'
|
| 120 |
+
- --context-parallel-size
|
| 121 |
+
- '1'
|
| 122 |
+
- --attention-backend
|
| 123 |
+
- flash
|
| 124 |
+
- --use-distributed-optimizer
|
| 125 |
+
- --overlap-grad-reduce
|
| 126 |
+
- --ddp-bucket-size
|
| 127 |
+
- '480000000'
|
| 128 |
+
- --ddp-pad-buckets-for-high-nccl-busbw
|
| 129 |
+
- --fused-residual-rmsnorm
|
| 130 |
+
- --tokenizer-type
|
| 131 |
+
- HuggingFaceTokenizer
|
| 132 |
+
- --tokenizer-model
|
| 133 |
+
- /mnt/shared-storage-gpfs2/plm-gpfs/public_data_model/model/Olmo-3-1025-7B-stage1-step1413814
|
| 134 |
+
- --vocab-size
|
| 135 |
+
- '100278'
|
| 136 |
+
- --data-cache-path
|
| 137 |
+
- /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/run/c22vq-dddiag256-full-0611113636/data_cache/real
|
| 138 |
+
- --split
|
| 139 |
+
- 100,0,0
|
| 140 |
+
- --no-create-attention-mask-in-dataloader
|
| 141 |
+
- --num-workers
|
| 142 |
+
- '2'
|
| 143 |
+
- --data-args-path
|
| 144 |
+
- /mnt/shared-storage-gpfs2/plm-gpfs/public_data_model/data/dolma3_mix-6T/megatron_olmo3_8192/data_args_path.txt
|
| 145 |
+
- --per-dataset-sequences-path
|
| 146 |
+
- /mnt/shared-storage-gpfs2/plm-gpfs/public_data_model/data/dolma3_mix-6T/megatron_olmo3_8192/per_dataset_sequences.json
|
| 147 |
+
- --no-mmap-bin-files
|
| 148 |
+
- --log-interval
|
| 149 |
+
- '1'
|
| 150 |
+
- --eval-iters
|
| 151 |
+
- '0'
|
| 152 |
+
- --eval-interval
|
| 153 |
+
- '1000'
|
| 154 |
+
- --rerun-mode
|
| 155 |
+
- disabled
|
| 156 |
+
- --log-throughput
|
| 157 |
+
- --log-timers-to-tensorboard
|
| 158 |
+
- --distributed-timeout-minutes
|
| 159 |
+
- '60'
|
| 160 |
+
- --tensorboard-dir
|
| 161 |
+
- /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/run/c22vq-dddiag256-full-0611113636/tensorboard
|
| 162 |
+
- --manual-gc
|
| 163 |
+
- --manual-gc-interval
|
| 164 |
+
- '10'
|
| 165 |
+
- --empty-unused-memory-level
|
| 166 |
+
- '0'
|
| 167 |
+
- --wandb-project
|
| 168 |
+
- ConceptLM
|
| 169 |
+
- --wandb-exp-name
|
| 170 |
+
- c22vq-dddiag256-full-0611113636
|
| 171 |
+
- --wandb-save-dir
|
| 172 |
+
- /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/run/c22vq-dddiag256-full-0611113636/wandb
|
| 173 |
+
- --wandb-entity
|
| 174 |
+
- iammi-nanjing-university
|
| 175 |
+
- --save
|
| 176 |
+
- /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/checkpoints
|
| 177 |
+
- --save-interval
|
| 178 |
+
- '10000'
|
| 179 |
+
- --ckpt-format
|
| 180 |
+
- torch_dist
|
| 181 |
+
- --conceptlm-backbone
|
| 182 |
+
- olmo3
|
| 183 |
+
- --conceptlm-encoder-layers
|
| 184 |
+
- '16'
|
| 185 |
+
- --conceptlm-decoder-layers
|
| 186 |
+
- '16'
|
| 187 |
+
- --conceptlm-special-layers
|
| 188 |
+
- '8'
|
| 189 |
+
- --conceptlm-chunk-size
|
| 190 |
+
- '4'
|
| 191 |
+
- --conceptlm-chunk-merge-method
|
| 192 |
+
- meanpooling
|
| 193 |
+
- --conceptlm-chunk-dualpath-alpha-init
|
| 194 |
+
- '0.5'
|
| 195 |
+
- --conceptlm-layer-norm-option
|
| 196 |
+
- normed_add
|
| 197 |
+
- --conceptlm-fusion-norm-alpha-init
|
| 198 |
+
- '0.1'
|
| 199 |
+
- --conceptlm-fusion-alpha-init
|
| 200 |
+
- '1.0'
|
| 201 |
+
- --conceptlm-hlm-attention-mode
|
| 202 |
+
- backbone_window
|
| 203 |
+
- --conceptlm-hlm-loss-weight
|
| 204 |
+
- '1.0'
|
| 205 |
+
- --conceptlm-vq-loss-weight
|
| 206 |
+
- '1.0'
|
| 207 |
+
- --conceptlm-v22-vq-codebook-size
|
| 208 |
+
- '128'
|
| 209 |
+
- --conceptlm-v22-vq-num-codebooks
|
| 210 |
+
- '32'
|
| 211 |
+
- --conceptlm-v22-vq-commitment-cost
|
| 212 |
+
- '0.25'
|
| 213 |
+
- --conceptlm-v22-vq-merge-mode
|
| 214 |
+
- raw_logits
|
| 215 |
+
- --conceptlm-v22-vq-hlm-loss-type
|
| 216 |
+
- mse
|
| 217 |
+
- --conceptlm-v22-vq-lr-mult
|
| 218 |
+
- '1.0'
|
| 219 |
+
- --conceptlm-v22-vq-highlevel-lr-mult
|
| 220 |
+
- '1.0'
|
| 221 |
+
- --conceptlm-v21-dd-two-route-add-concept-source
|
| 222 |
+
- final
|
| 223 |
+
- --conceptlm-v21-dd-two-route-add-beta-init
|
| 224 |
+
- '0.05'
|
| 225 |
+
- --conceptlm-v21-dd-two-route-add-every-n-layers
|
| 226 |
+
- '1'
|
| 227 |
+
- --conceptlm-v21-dd-two-route-add-concept-route-first-n
|
| 228 |
+
- '-1'
|
| 229 |
+
- --conceptlm-v21-dd-two-route-add-decoder-hidden-size
|
| 230 |
+
- '0'
|
| 231 |
+
- --conceptlm-v21-dd-two-route-add-concept-hidden-size
|
| 232 |
+
- '0'
|
| 233 |
+
- --conceptlm-v21-dd-encoder-self-dd-every-n-layers
|
| 234 |
+
- '1'
|
| 235 |
+
- --conceptlm-v21-dd-encoder-self-dd-hidden-size
|
| 236 |
+
- '0'
|
| 237 |
+
- --conceptlm-v21-dd-concept-self-dd-every-n-layers
|
| 238 |
+
- '1'
|
| 239 |
+
- --conceptlm-v21-dd-concept-self-dd-hidden-size
|
| 240 |
+
- '0'
|
| 241 |
+
- --conceptlm-v21-residual-flow-beta-init
|
| 242 |
+
- '0.02'
|
| 243 |
+
- --conceptlm-v21-residual-flow-route-hidden-size
|
| 244 |
+
- '0'
|
| 245 |
+
- --conceptlm-v21-final-read-concept-gate-init-final
|
| 246 |
+
- '0.5'
|
| 247 |
+
- --conceptlm-v21-final-read-concept-gate-target-final
|
| 248 |
+
- '0.5'
|
| 249 |
+
- --conceptlm-v21-final-read-concept-gate-reg-weight
|
| 250 |
+
- '0.0'
|
| 251 |
+
- --conceptlm-shift-feature
|
| 252 |
+
- --no-conceptlm-enable-chunk-dualpath-smoothing
|
| 253 |
+
- --conceptlm-v22-vq-detach-hlm-target
|
| 254 |
+
- --conceptlm-v21-dd-two-route-add
|
| 255 |
+
- --no-conceptlm-v21-dd-two-route-add-enable-raw-concept-route
|
| 256 |
+
- --conceptlm-v21-dd-two-route-add-enable-final-concept-route
|
| 257 |
+
- --no-conceptlm-v21-dd-two-route-add-use-softmax
|
| 258 |
+
- --no-conceptlm-v21-dd-two-route-add-disable-decoder-dd
|
| 259 |
+
- --no-conceptlm-v21-dd-two-route-add-decoder-use-layernorm
|
| 260 |
+
- --conceptlm-v21-dd-two-route-add-decoder-use-softmax
|
| 261 |
+
- --conceptlm-v21-dd-two-route-add-concept-use-layernorm
|
| 262 |
+
- --conceptlm-v21-dd-encoder-self-dd
|
| 263 |
+
- --no-conceptlm-v21-dd-encoder-self-dd-use-layernorm
|
| 264 |
+
- --conceptlm-v21-dd-concept-self-dd
|
| 265 |
+
- --no-conceptlm-v21-dd-concept-self-dd-use-layernorm
|
| 266 |
+
- --no-conceptlm-v21-enable-full-residual-flow
|
| 267 |
+
- --conceptlm-v21-enable-concept-read-encoder
|
| 268 |
+
- --conceptlm-v21-enable-decoder-read-encoder
|
| 269 |
+
- --conceptlm-v21-enable-decoder-read-concept
|
| 270 |
+
- --conceptlm-v21-residual-flow-route-use-softmax
|
| 271 |
+
- --conceptlm-v21-residual-flow-source-use-layernorm
|
| 272 |
+
- --conceptlm-v21-residual-flow-shared-source-norm
|
| 273 |
+
- --conceptlm-v21-final-read-concept-gate
|
| 274 |
+
- --conceptlm-v21-compile-residual-flow-routes
|
| 275 |
+
- --conceptlm-v21-compile-dd-routes
|
| 276 |
+
program: /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/repo/pretrain_conceptlm_v22_vq.py
|
| 277 |
+
code_path: repo/pretrain_conceptlm_v22_vq.py
|
| 278 |
+
code_path_local: pretrain_conceptlm_v22_vq.py
|
| 279 |
+
git:
|
| 280 |
+
remote_url: https://github.com/Liu-yuliang/train_conceptlm_V22_olmo_7B.git
|
| 281 |
+
commit: 1f85baff62a930b57427302e65f7822c0cd8b3a8
|
| 282 |
+
root: /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/run/c22vq-dddiag256-full-0611113636/wandb
|
| 283 |
+
host: gpu-l-lg-cmc-h-h200-0011.host.h.pjlab.org.cn
|
| 284 |
+
executable: /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/miniconda3/envs/megatron-h200-cu128-py312/bin/python
|
| 285 |
+
cpu_count: 96
|
| 286 |
+
cpu_count_logical: 96
|
| 287 |
+
gpu_type: NVIDIA H200
|
| 288 |
+
gpu_count: 8
|
| 289 |
+
disk:
|
| 290 |
+
/:
|
| 291 |
+
total: '3838880616448'
|
| 292 |
+
used: '181383819264'
|
| 293 |
+
memory:
|
| 294 |
+
total: '824633720832'
|
| 295 |
+
gpu_nvidia:
|
| 296 |
+
- name: NVIDIA H200
|
| 297 |
+
memory_total: '150754820096'
|
| 298 |
+
cuda_cores: 16896
|
| 299 |
+
architecture: Hopper
|
| 300 |
+
uuid: GPU-1c3a1fc6-2417-649c-84dd-ead22666f5a3
|
| 301 |
+
- name: NVIDIA H200
|
| 302 |
+
memory_total: '150754820096'
|
| 303 |
+
cuda_cores: 16896
|
| 304 |
+
architecture: Hopper
|
| 305 |
+
uuid: GPU-0256efa9-df8e-a893-f15d-db243c6be812
|
| 306 |
+
- name: NVIDIA H200
|
| 307 |
+
memory_total: '150754820096'
|
| 308 |
+
cuda_cores: 16896
|
| 309 |
+
architecture: Hopper
|
| 310 |
+
uuid: GPU-c0936028-dda8-a572-e197-94ad7bc69c25
|
| 311 |
+
- name: NVIDIA H200
|
| 312 |
+
memory_total: '150754820096'
|
| 313 |
+
cuda_cores: 16896
|
| 314 |
+
architecture: Hopper
|
| 315 |
+
uuid: GPU-7627700d-ae93-b6d4-16cf-c87166a980ee
|
| 316 |
+
- name: NVIDIA H200
|
| 317 |
+
memory_total: '150754820096'
|
| 318 |
+
cuda_cores: 16896
|
| 319 |
+
architecture: Hopper
|
| 320 |
+
uuid: GPU-0d5ed50b-33e1-5228-6efc-5e258a184e48
|
| 321 |
+
- name: NVIDIA H200
|
| 322 |
+
memory_total: '150754820096'
|
| 323 |
+
cuda_cores: 16896
|
| 324 |
+
architecture: Hopper
|
| 325 |
+
uuid: GPU-e5b93d5a-3d32-8cb8-f8fe-8d0ca532d491
|
| 326 |
+
- name: NVIDIA H200
|
| 327 |
+
memory_total: '150754820096'
|
| 328 |
+
cuda_cores: 16896
|
| 329 |
+
architecture: Hopper
|
| 330 |
+
uuid: GPU-29ad5514-f922-e820-4058-9d5eab0ac196
|
| 331 |
+
- name: NVIDIA H200
|
| 332 |
+
memory_total: '150754820096'
|
| 333 |
+
cuda_cores: 16896
|
| 334 |
+
architecture: Hopper
|
| 335 |
+
uuid: GPU-b70a6b7c-bc79-07c5-56f9-e24c435544be
|
| 336 |
+
cuda_version: '12.8'
|
| 337 |
+
writer_id: lwpdwcz10y13ud608inuzplox070dihm
|
| 338 |
+
tensor_model_parallel_size:
|
| 339 |
+
desc: null
|
| 340 |
+
value: 1
|
| 341 |
+
pipeline_model_parallel_comm_backend:
|
| 342 |
+
desc: null
|
| 343 |
+
value: null
|
| 344 |
+
pipeline_model_parallel_size:
|
| 345 |
+
desc: null
|
| 346 |
+
value: 1
|
| 347 |
+
sequence_parallel:
|
| 348 |
+
desc: null
|
| 349 |
+
value: false
|
| 350 |
+
context_parallel_size:
|
| 351 |
+
desc: null
|
| 352 |
+
value: 1
|
| 353 |
+
hierarchical_context_parallel_sizes:
|
| 354 |
+
desc: null
|
| 355 |
+
value: null
|
| 356 |
+
max_seqlen_per_dp_cp_rank:
|
| 357 |
+
desc: null
|
| 358 |
+
value: null
|
| 359 |
+
hybrid_context_parallel:
|
| 360 |
+
desc: null
|
| 361 |
+
value: false
|
| 362 |
+
expert_model_parallel_size:
|
| 363 |
+
desc: null
|
| 364 |
+
value: 1
|
| 365 |
+
expert_tensor_parallel_size:
|
| 366 |
+
desc: null
|
| 367 |
+
value: 1
|
| 368 |
+
perform_initialization:
|
| 369 |
+
desc: null
|
| 370 |
+
value: true
|
| 371 |
+
use_cpu_initialization:
|
| 372 |
+
desc: null
|
| 373 |
+
value: null
|
| 374 |
+
fp16:
|
| 375 |
+
desc: null
|
| 376 |
+
value: false
|
| 377 |
+
bf16:
|
| 378 |
+
desc: null
|
| 379 |
+
value: true
|
| 380 |
+
deterministic_mode:
|
| 381 |
+
desc: null
|
| 382 |
+
value: false
|
| 383 |
+
tp_comm_overlap:
|
| 384 |
+
desc: null
|
| 385 |
+
value: false
|
| 386 |
+
tp_comm_bulk_wgrad:
|
| 387 |
+
desc: null
|
| 388 |
+
value: true
|
| 389 |
+
tp_comm_bulk_dgrad:
|
| 390 |
+
desc: null
|
| 391 |
+
value: true
|
| 392 |
+
tp_comm_overlap_ag:
|
| 393 |
+
desc: null
|
| 394 |
+
value: true
|
| 395 |
+
tp_comm_overlap_rs:
|
| 396 |
+
desc: null
|
| 397 |
+
value: true
|
| 398 |
+
tp_comm_overlap_rs_dgrad:
|
| 399 |
+
desc: null
|
| 400 |
+
value: false
|
| 401 |
+
tp_comm_split_ag:
|
| 402 |
+
desc: null
|
| 403 |
+
value: true
|
| 404 |
+
tp_comm_split_rs:
|
| 405 |
+
desc: null
|
| 406 |
+
value: true
|
| 407 |
+
cross_entropy_loss_fusion:
|
| 408 |
+
desc: null
|
| 409 |
+
value: true
|
| 410 |
+
cross_entropy_fusion_impl:
|
| 411 |
+
desc: null
|
| 412 |
+
value: native
|
| 413 |
+
tp_comm_bootstrap_backend:
|
| 414 |
+
desc: null
|
| 415 |
+
value: nccl
|
| 416 |
+
overlap_moe_expert_parallel_comm:
|
| 417 |
+
desc: null
|
| 418 |
+
value: false
|
| 419 |
+
delay_wgrad_compute:
|
| 420 |
+
desc: null
|
| 421 |
+
value: false
|
| 422 |
+
overlap_dispatch_backward_with_experts_wgrad:
|
| 423 |
+
desc: null
|
| 424 |
+
value: false
|
| 425 |
+
ep_overlap_early_attn_memory_release:
|
| 426 |
+
desc: null
|
| 427 |
+
value: false
|
| 428 |
+
use_ring_exchange_p2p:
|
| 429 |
+
desc: null
|
| 430 |
+
value: false
|
| 431 |
+
defer_embedding_wgrad_compute:
|
| 432 |
+
desc: null
|
| 433 |
+
value: false
|
| 434 |
+
wgrad_deferral_limit:
|
| 435 |
+
desc: null
|
| 436 |
+
value: 0
|
| 437 |
+
overlap_p2p_comm_warmup_flush:
|
| 438 |
+
desc: null
|
| 439 |
+
value: false
|
| 440 |
+
microbatch_group_size_per_vp_stage:
|
| 441 |
+
desc: null
|
| 442 |
+
value: null
|
| 443 |
+
mtp_standalone:
|
| 444 |
+
desc: null
|
| 445 |
+
value: false
|
| 446 |
+
cpu_offloading_num_layers:
|
| 447 |
+
desc: null
|
| 448 |
+
value: 0
|
| 449 |
+
cpu_offloading_retain_pinned_cpu_buffers:
|
| 450 |
+
desc: null
|
| 451 |
+
value: false
|
| 452 |
+
num_layers:
|
| 453 |
+
desc: null
|
| 454 |
+
value: 32
|
| 455 |
+
mtp_num_layers:
|
| 456 |
+
desc: null
|
| 457 |
+
value: null
|
| 458 |
+
mtp_loss_scaling_factor:
|
| 459 |
+
desc: null
|
| 460 |
+
value: 0.1
|
| 461 |
+
mtp_use_repeated_layer:
|
| 462 |
+
desc: null
|
| 463 |
+
value: false
|
| 464 |
+
mtp_hybrid_override_pattern:
|
| 465 |
+
desc: null
|
| 466 |
+
value: null
|
| 467 |
+
account_for_embedding_in_pipeline_split:
|
| 468 |
+
desc: null
|
| 469 |
+
value: false
|
| 470 |
+
account_for_loss_in_pipeline_split:
|
| 471 |
+
desc: null
|
| 472 |
+
value: false
|
| 473 |
+
hidden_size:
|
| 474 |
+
desc: null
|
| 475 |
+
value: 4096
|
| 476 |
+
num_attention_heads:
|
| 477 |
+
desc: null
|
| 478 |
+
value: 32
|
| 479 |
+
attention_backend:
|
| 480 |
+
desc: null
|
| 481 |
+
value: flash
|
| 482 |
+
softmax_type:
|
| 483 |
+
desc: null
|
| 484 |
+
value: vanilla
|
| 485 |
+
num_query_groups:
|
| 486 |
+
desc: null
|
| 487 |
+
value: 1
|
| 488 |
+
ffn_hidden_size:
|
| 489 |
+
desc: null
|
| 490 |
+
value: 11008
|
| 491 |
+
kv_channels:
|
| 492 |
+
desc: null
|
| 493 |
+
value: 128
|
| 494 |
+
hidden_dropout:
|
| 495 |
+
desc: null
|
| 496 |
+
value: 0.0
|
| 497 |
+
attention_dropout:
|
| 498 |
+
desc: null
|
| 499 |
+
value: 0.0
|
| 500 |
+
fp32_residual_connection:
|
| 501 |
+
desc: null
|
| 502 |
+
value: false
|
| 503 |
+
apply_residual_connection_post_layernorm:
|
| 504 |
+
desc: null
|
| 505 |
+
value: false
|
| 506 |
+
layernorm_epsilon:
|
| 507 |
+
desc: null
|
| 508 |
+
value: 1.0e-06
|
| 509 |
+
layernorm_zero_centered_gamma:
|
| 510 |
+
desc: null
|
| 511 |
+
value: false
|
| 512 |
+
add_bias_linear:
|
| 513 |
+
desc: null
|
| 514 |
+
value: false
|
| 515 |
+
add_qkv_bias:
|
| 516 |
+
desc: null
|
| 517 |
+
value: false
|
| 518 |
+
glu_linear_offset:
|
| 519 |
+
desc: null
|
| 520 |
+
value: 0.0
|
| 521 |
+
activation_func_clamp_value:
|
| 522 |
+
desc: null
|
| 523 |
+
value: null
|
| 524 |
+
rotary_interleaved:
|
| 525 |
+
desc: null
|
| 526 |
+
value: false
|
| 527 |
+
normalization:
|
| 528 |
+
desc: null
|
| 529 |
+
value: RMSNorm
|
| 530 |
+
qk_layernorm:
|
| 531 |
+
desc: null
|
| 532 |
+
value: true
|
| 533 |
+
qk_l2_norm:
|
| 534 |
+
desc: null
|
| 535 |
+
value: false
|
| 536 |
+
qk_clip:
|
| 537 |
+
desc: null
|
| 538 |
+
value: false
|
| 539 |
+
qk_clip_alpha:
|
| 540 |
+
desc: null
|
| 541 |
+
value: 0.5
|
| 542 |
+
qk_clip_threshold:
|
| 543 |
+
desc: null
|
| 544 |
+
value: 100
|
| 545 |
+
attention_output_gate:
|
| 546 |
+
desc: null
|
| 547 |
+
value: false
|
| 548 |
+
calculate_per_token_loss:
|
| 549 |
+
desc: null
|
| 550 |
+
value: false
|
| 551 |
+
multi_latent_attention:
|
| 552 |
+
desc: null
|
| 553 |
+
value: false
|
| 554 |
+
experimental_attention_variant:
|
| 555 |
+
desc: null
|
| 556 |
+
value: null
|
| 557 |
+
dsa_indexer_n_heads:
|
| 558 |
+
desc: null
|
| 559 |
+
value: null
|
| 560 |
+
dsa_indexer_head_dim:
|
| 561 |
+
desc: null
|
| 562 |
+
value: null
|
| 563 |
+
dsa_indexer_topk:
|
| 564 |
+
desc: null
|
| 565 |
+
value: null
|
| 566 |
+
dsa_indexer_loss_coeff:
|
| 567 |
+
desc: null
|
| 568 |
+
value: null
|
| 569 |
+
dsa_indexer_use_sparse_loss:
|
| 570 |
+
desc: null
|
| 571 |
+
value: false
|
| 572 |
+
linear_conv_kernel_dim:
|
| 573 |
+
desc: null
|
| 574 |
+
value: 4
|
| 575 |
+
linear_key_head_dim:
|
| 576 |
+
desc: null
|
| 577 |
+
value: 128
|
| 578 |
+
linear_value_head_dim:
|
| 579 |
+
desc: null
|
| 580 |
+
value: 128
|
| 581 |
+
linear_num_key_heads:
|
| 582 |
+
desc: null
|
| 583 |
+
value: 16
|
| 584 |
+
linear_num_value_heads:
|
| 585 |
+
desc: null
|
| 586 |
+
value: 32
|
| 587 |
+
init_method_std:
|
| 588 |
+
desc: null
|
| 589 |
+
value: 0.02
|
| 590 |
+
init_method_variant:
|
| 591 |
+
desc: null
|
| 592 |
+
value: scaled_truncated_normal
|
| 593 |
+
embedding_init_method_std:
|
| 594 |
+
desc: null
|
| 595 |
+
value: null
|
| 596 |
+
init_model_with_meta_device:
|
| 597 |
+
desc: null
|
| 598 |
+
value: false
|
| 599 |
+
use_mup:
|
| 600 |
+
desc: null
|
| 601 |
+
value: false
|
| 602 |
+
mup_width_mult:
|
| 603 |
+
desc: null
|
| 604 |
+
value: 1.0
|
| 605 |
+
mup_base_hidden_size:
|
| 606 |
+
desc: null
|
| 607 |
+
value: null
|
| 608 |
+
mup_embedding_mult:
|
| 609 |
+
desc: null
|
| 610 |
+
value: 1.0
|
| 611 |
+
mup_output_mult:
|
| 612 |
+
desc: null
|
| 613 |
+
value: 1.0
|
| 614 |
+
mup_base_head_dim:
|
| 615 |
+
desc: null
|
| 616 |
+
value: null
|
| 617 |
+
mup_attn_scale_power:
|
| 618 |
+
desc: null
|
| 619 |
+
value: 1.0
|
| 620 |
+
apply_query_key_layer_scaling:
|
| 621 |
+
desc: null
|
| 622 |
+
value: false
|
| 623 |
+
disable_bf16_reduced_precision_matmul:
|
| 624 |
+
desc: null
|
| 625 |
+
value: false
|
| 626 |
+
use_fused_weighted_squared_relu:
|
| 627 |
+
desc: null
|
| 628 |
+
value: false
|
| 629 |
+
fused_residual_rmsnorm:
|
| 630 |
+
desc: null
|
| 631 |
+
value: true
|
| 632 |
+
recompute_granularity:
|
| 633 |
+
desc: null
|
| 634 |
+
value: null
|
| 635 |
+
recompute_method:
|
| 636 |
+
desc: null
|
| 637 |
+
value: null
|
| 638 |
+
recompute_num_layers:
|
| 639 |
+
desc: null
|
| 640 |
+
value: null
|
| 641 |
+
distribute_saved_activations:
|
| 642 |
+
desc: null
|
| 643 |
+
value: false
|
| 644 |
+
recompute_modules:
|
| 645 |
+
desc: null
|
| 646 |
+
value: null
|
| 647 |
+
fp8:
|
| 648 |
+
desc: null
|
| 649 |
+
value: null
|
| 650 |
+
fp8_recipe:
|
| 651 |
+
desc: null
|
| 652 |
+
value: delayed
|
| 653 |
+
fp8_quantizer_factory:
|
| 654 |
+
desc: null
|
| 655 |
+
value: null
|
| 656 |
+
fp8_margin:
|
| 657 |
+
desc: null
|
| 658 |
+
value: 0
|
| 659 |
+
fp8_interval:
|
| 660 |
+
desc: null
|
| 661 |
+
value: 1
|
| 662 |
+
fp8_amax_history_len:
|
| 663 |
+
desc: null
|
| 664 |
+
value: 1
|
| 665 |
+
fp8_amax_compute_algo:
|
| 666 |
+
desc: null
|
| 667 |
+
value: most_recent
|
| 668 |
+
fp8_wgrad:
|
| 669 |
+
desc: null
|
| 670 |
+
value: true
|
| 671 |
+
first_last_layers_bf16:
|
| 672 |
+
desc: null
|
| 673 |
+
value: false
|
| 674 |
+
num_layers_at_start_in_bf16:
|
| 675 |
+
desc: null
|
| 676 |
+
value: 1
|
| 677 |
+
num_layers_at_end_in_bf16:
|
| 678 |
+
desc: null
|
| 679 |
+
value: 1
|
| 680 |
+
use_kitchen_attention:
|
| 681 |
+
desc: null
|
| 682 |
+
value: false
|
| 683 |
+
kitchen_attention_backend:
|
| 684 |
+
desc: null
|
| 685 |
+
value: sdpa
|
| 686 |
+
fp4:
|
| 687 |
+
desc: null
|
| 688 |
+
value: null
|
| 689 |
+
fp4_recipe:
|
| 690 |
+
desc: null
|
| 691 |
+
value: nvfp4
|
| 692 |
+
fp4_quantizer_factory:
|
| 693 |
+
desc: null
|
| 694 |
+
value: null
|
| 695 |
+
moe_shared_expert_intermediate_size:
|
| 696 |
+
desc: null
|
| 697 |
+
value: null
|
| 698 |
+
moe_shared_expert_gate:
|
| 699 |
+
desc: null
|
| 700 |
+
value: false
|
| 701 |
+
moe_shared_expert_overlap:
|
| 702 |
+
desc: null
|
| 703 |
+
value: false
|
| 704 |
+
moe_ffn_hidden_size:
|
| 705 |
+
desc: null
|
| 706 |
+
value: null
|
| 707 |
+
moe_router_topk:
|
| 708 |
+
desc: null
|
| 709 |
+
value: 2
|
| 710 |
+
moe_enable_routing_replay:
|
| 711 |
+
desc: null
|
| 712 |
+
value: false
|
| 713 |
+
moe_router_padding_for_quantization:
|
| 714 |
+
desc: null
|
| 715 |
+
value: false
|
| 716 |
+
moe_router_padding_for_fp8:
|
| 717 |
+
desc: null
|
| 718 |
+
value: false
|
| 719 |
+
moe_router_num_groups:
|
| 720 |
+
desc: null
|
| 721 |
+
value: null
|
| 722 |
+
moe_router_group_topk:
|
| 723 |
+
desc: null
|
| 724 |
+
value: null
|
| 725 |
+
moe_router_pre_softmax:
|
| 726 |
+
desc: null
|
| 727 |
+
value: false
|
| 728 |
+
moe_router_topk_scaling_factor:
|
| 729 |
+
desc: null
|
| 730 |
+
value: null
|
| 731 |
+
moe_router_score_function:
|
| 732 |
+
desc: null
|
| 733 |
+
value: softmax
|
| 734 |
+
moe_router_dtype:
|
| 735 |
+
desc: null
|
| 736 |
+
value: null
|
| 737 |
+
moe_router_enable_expert_bias:
|
| 738 |
+
desc: null
|
| 739 |
+
value: false
|
| 740 |
+
moe_router_bias_update_rate:
|
| 741 |
+
desc: null
|
| 742 |
+
value: 0.001
|
| 743 |
+
moe_router_force_load_balancing:
|
| 744 |
+
desc: null
|
| 745 |
+
value: false
|
| 746 |
+
moe_router_force_biased:
|
| 747 |
+
desc: null
|
| 748 |
+
value: null
|
| 749 |
+
moe_grouped_gemm:
|
| 750 |
+
desc: null
|
| 751 |
+
value: false
|
| 752 |
+
moe_z_loss_coeff:
|
| 753 |
+
desc: null
|
| 754 |
+
value: null
|
| 755 |
+
moe_input_jitter_eps:
|
| 756 |
+
desc: null
|
| 757 |
+
value: null
|
| 758 |
+
moe_token_dispatcher_type:
|
| 759 |
+
desc: null
|
| 760 |
+
value: allgather
|
| 761 |
+
moe_enable_deepep:
|
| 762 |
+
desc: null
|
| 763 |
+
value: false
|
| 764 |
+
moe_flex_dispatcher_backend:
|
| 765 |
+
desc: null
|
| 766 |
+
value: deepep
|
| 767 |
+
moe_permute_fusion_into_hybridep:
|
| 768 |
+
desc: null
|
| 769 |
+
value: false
|
| 770 |
+
moe_per_layer_logging:
|
| 771 |
+
desc: null
|
| 772 |
+
value: false
|
| 773 |
+
moe_expert_capacity_factor:
|
| 774 |
+
desc: null
|
| 775 |
+
value: null
|
| 776 |
+
moe_pad_expert_input_to_capacity:
|
| 777 |
+
desc: null
|
| 778 |
+
value: false
|
| 779 |
+
moe_pad_experts_for_cuda_graph_inference:
|
| 780 |
+
desc: null
|
| 781 |
+
value: false
|
| 782 |
+
moe_token_drop_policy:
|
| 783 |
+
desc: null
|
| 784 |
+
value: probs
|
| 785 |
+
moe_layer_recompute:
|
| 786 |
+
desc: null
|
| 787 |
+
value: false
|
| 788 |
+
moe_permute_fusion:
|
| 789 |
+
desc: null
|
| 790 |
+
value: false
|
| 791 |
+
moe_router_fusion:
|
| 792 |
+
desc: null
|
| 793 |
+
value: false
|
| 794 |
+
moe_apply_probs_on_input:
|
| 795 |
+
desc: null
|
| 796 |
+
value: false
|
| 797 |
+
moe_latent_size:
|
| 798 |
+
desc: null
|
| 799 |
+
value: null
|
| 800 |
+
moe_deepep_num_sms:
|
| 801 |
+
desc: null
|
| 802 |
+
value: 20
|
| 803 |
+
moe_hybridep_num_sms:
|
| 804 |
+
desc: null
|
| 805 |
+
value: null
|
| 806 |
+
moe_hybridep_num_blocks_permute:
|
| 807 |
+
desc: null
|
| 808 |
+
value: null
|
| 809 |
+
moe_hybridep_num_blocks_unpermute:
|
| 810 |
+
desc: null
|
| 811 |
+
value: null
|
| 812 |
+
enable_cuda_graph:
|
| 813 |
+
desc: null
|
| 814 |
+
value: false
|
| 815 |
+
cuda_graph_warmup_steps:
|
| 816 |
+
desc: null
|
| 817 |
+
value: 3
|
| 818 |
+
external_cuda_graph:
|
| 819 |
+
desc: null
|
| 820 |
+
value: false
|
| 821 |
+
cuda_graph_impl:
|
| 822 |
+
desc: null
|
| 823 |
+
value: none
|
| 824 |
+
clone_scatter_output_in_embedding:
|
| 825 |
+
desc: null
|
| 826 |
+
value: true
|
| 827 |
+
config_logger_dir:
|
| 828 |
+
desc: null
|
| 829 |
+
value: ''
|
| 830 |
+
flash_decode:
|
| 831 |
+
desc: null
|
| 832 |
+
value: false
|
| 833 |
+
batch_invariant_mode:
|
| 834 |
+
desc: null
|
| 835 |
+
value: false
|
| 836 |
+
use_te_activation_func:
|
| 837 |
+
desc: null
|
| 838 |
+
value: false
|
| 839 |
+
symmetric_ar_type:
|
| 840 |
+
desc: null
|
| 841 |
+
value: null
|
| 842 |
+
nccl_all_reduce_for_prefill:
|
| 843 |
+
desc: null
|
| 844 |
+
value: false
|
| 845 |
+
inference_fuse_tp_communication:
|
| 846 |
+
desc: null
|
| 847 |
+
value: false
|
| 848 |
+
inference_disable_triton_nvls_kernels:
|
| 849 |
+
desc: null
|
| 850 |
+
value: false
|
| 851 |
+
inference_grouped_gemm_backend:
|
| 852 |
+
desc: null
|
| 853 |
+
value: vllm
|
| 854 |
+
inference_moe_disable_fused_quant_kernels:
|
| 855 |
+
desc: null
|
| 856 |
+
value: false
|
| 857 |
+
inference_moe_token_dispatcher_type:
|
| 858 |
+
desc: null
|
| 859 |
+
value: nvls
|
| 860 |
+
mrope_section:
|
| 861 |
+
desc: null
|
| 862 |
+
value: null
|
| 863 |
+
is_hybrid_model:
|
| 864 |
+
desc: null
|
| 865 |
+
value: false
|
| 866 |
+
mamba_state_dim:
|
| 867 |
+
desc: null
|
| 868 |
+
value: 128
|
| 869 |
+
mamba_head_dim:
|
| 870 |
+
desc: null
|
| 871 |
+
value: 64
|
| 872 |
+
mamba_num_groups:
|
| 873 |
+
desc: null
|
| 874 |
+
value: 8
|
| 875 |
+
mamba_num_heads:
|
| 876 |
+
desc: null
|
| 877 |
+
value: null
|
| 878 |
+
use_mamba_mem_eff_path:
|
| 879 |
+
desc: null
|
| 880 |
+
value: true
|
| 881 |
+
mlp_chunks_for_prefill:
|
| 882 |
+
desc: null
|
| 883 |
+
value: 1
|
| 884 |
+
mlp_chunks_for_training:
|
| 885 |
+
desc: null
|
| 886 |
+
value: 1
|
| 887 |
+
transformer_impl:
|
| 888 |
+
desc: null
|
| 889 |
+
value: transformer_engine
|
| 890 |
+
fine_grained_activation_offloading:
|
| 891 |
+
desc: null
|
| 892 |
+
value: false
|
| 893 |
+
offload_modules:
|
| 894 |
+
desc: null
|
| 895 |
+
value: []
|
| 896 |
+
min_offloaded_tensor_size:
|
| 897 |
+
desc: null
|
| 898 |
+
value: 1048576
|
| 899 |
+
encoder_num_layers:
|
| 900 |
+
desc: null
|
| 901 |
+
value: 32
|
| 902 |
+
decoder_num_layers:
|
| 903 |
+
desc: null
|
| 904 |
+
value: null
|
| 905 |
+
group_query_attention:
|
| 906 |
+
desc: null
|
| 907 |
+
value: false
|
| 908 |
+
window_size:
|
| 909 |
+
desc: null
|
| 910 |
+
value:
|
| 911 |
+
- 4096
|
| 912 |
+
- 0
|
| 913 |
+
window_attn_skip_freq:
|
| 914 |
+
desc: null
|
| 915 |
+
value: 4
|
| 916 |
+
max_position_embeddings:
|
| 917 |
+
desc: null
|
| 918 |
+
value: 8192
|
| 919 |
+
position_embedding_type:
|
| 920 |
+
desc: null
|
| 921 |
+
value: rope
|
| 922 |
+
relative_attention_num_buckets:
|
| 923 |
+
desc: null
|
| 924 |
+
value: 32
|
| 925 |
+
relative_attention_max_distance:
|
| 926 |
+
desc: null
|
| 927 |
+
value: 128
|
| 928 |
+
use_rotary_position_embeddings:
|
| 929 |
+
desc: null
|
| 930 |
+
value: false
|
| 931 |
+
rotary_base:
|
| 932 |
+
desc: null
|
| 933 |
+
value: 500000
|
| 934 |
+
rotary_percent:
|
| 935 |
+
desc: null
|
| 936 |
+
value: 1.0
|
| 937 |
+
rotary_seq_len_interpolation_factor:
|
| 938 |
+
desc: null
|
| 939 |
+
value: null
|
| 940 |
+
use_rope_scaling:
|
| 941 |
+
desc: null
|
| 942 |
+
value: false
|
| 943 |
+
rope_scaling_factor:
|
| 944 |
+
desc: null
|
| 945 |
+
value: 8.0
|
| 946 |
+
no_rope_freq:
|
| 947 |
+
desc: null
|
| 948 |
+
value: null
|
| 949 |
+
add_position_embedding:
|
| 950 |
+
desc: null
|
| 951 |
+
value: true
|
| 952 |
+
make_vocab_size_divisible_by:
|
| 953 |
+
desc: null
|
| 954 |
+
value: 1
|
| 955 |
+
openai_gelu:
|
| 956 |
+
desc: null
|
| 957 |
+
value: false
|
| 958 |
+
squared_relu:
|
| 959 |
+
desc: null
|
| 960 |
+
value: false
|
| 961 |
+
swiglu:
|
| 962 |
+
desc: null
|
| 963 |
+
value: true
|
| 964 |
+
quick_geglu:
|
| 965 |
+
desc: null
|
| 966 |
+
value: false
|
| 967 |
+
onnx_safe:
|
| 968 |
+
desc: null
|
| 969 |
+
value: null
|
| 970 |
+
bert_binary_head:
|
| 971 |
+
desc: null
|
| 972 |
+
value: true
|
| 973 |
+
untie_embeddings_and_output_weights:
|
| 974 |
+
desc: null
|
| 975 |
+
value: true
|
| 976 |
+
weight_decay:
|
| 977 |
+
desc: null
|
| 978 |
+
value: 0.1
|
| 979 |
+
apply_wd_to_qk_layernorm:
|
| 980 |
+
desc: null
|
| 981 |
+
value: false
|
| 982 |
+
clip_grad:
|
| 983 |
+
desc: null
|
| 984 |
+
value: 1.0
|
| 985 |
+
adam_beta1:
|
| 986 |
+
desc: null
|
| 987 |
+
value: 0.9
|
| 988 |
+
adam_beta2:
|
| 989 |
+
desc: null
|
| 990 |
+
value: 0.95
|
| 991 |
+
adam_eps:
|
| 992 |
+
desc: null
|
| 993 |
+
value: 1.0e-08
|
| 994 |
+
sgd_momentum:
|
| 995 |
+
desc: null
|
| 996 |
+
value: 0.9
|
| 997 |
+
muon_momentum:
|
| 998 |
+
desc: null
|
| 999 |
+
value: 0.9
|
| 1000 |
+
muon_split_qkv:
|
| 1001 |
+
desc: null
|
| 1002 |
+
value: true
|
| 1003 |
+
muon_nesterov:
|
| 1004 |
+
desc: null
|
| 1005 |
+
value: false
|
| 1006 |
+
muon_scale_mode:
|
| 1007 |
+
desc: null
|
| 1008 |
+
value: spectral
|
| 1009 |
+
muon_fp32_matmul_prec:
|
| 1010 |
+
desc: null
|
| 1011 |
+
value: medium
|
| 1012 |
+
muon_coefficient_type:
|
| 1013 |
+
desc: null
|
| 1014 |
+
value: quintic
|
| 1015 |
+
muon_num_ns_steps:
|
| 1016 |
+
desc: null
|
| 1017 |
+
value: 5
|
| 1018 |
+
muon_tp_mode:
|
| 1019 |
+
desc: null
|
| 1020 |
+
value: blockwise
|
| 1021 |
+
muon_extra_scale_factor:
|
| 1022 |
+
desc: null
|
| 1023 |
+
value: 1.0
|
| 1024 |
+
muon_scalar_optimizer:
|
| 1025 |
+
desc: null
|
| 1026 |
+
value: adam
|
| 1027 |
+
lion_beta1:
|
| 1028 |
+
desc: null
|
| 1029 |
+
value: 0.95
|
| 1030 |
+
lion_beta2:
|
| 1031 |
+
desc: null
|
| 1032 |
+
value: 0.98
|
| 1033 |
+
no_weight_decay_cond_type:
|
| 1034 |
+
desc: null
|
| 1035 |
+
value: null
|
| 1036 |
+
profile:
|
| 1037 |
+
desc: null
|
| 1038 |
+
value: false
|
| 1039 |
+
profile_step_start:
|
| 1040 |
+
desc: null
|
| 1041 |
+
value: 10
|
| 1042 |
+
profile_step_end:
|
| 1043 |
+
desc: null
|
| 1044 |
+
value: 12
|
| 1045 |
+
use_pytorch_profiler:
|
| 1046 |
+
desc: null
|
| 1047 |
+
value: false
|
| 1048 |
+
pytorch_profiler_collect_shapes:
|
| 1049 |
+
desc: null
|
| 1050 |
+
value: false
|
| 1051 |
+
pytorch_profiler_collect_callstack:
|
| 1052 |
+
desc: null
|
| 1053 |
+
value: false
|
| 1054 |
+
pytorch_profiler_collect_chakra:
|
| 1055 |
+
desc: null
|
| 1056 |
+
value: false
|
| 1057 |
+
profile_ranks:
|
| 1058 |
+
desc: null
|
| 1059 |
+
value: []
|
| 1060 |
+
record_memory_history:
|
| 1061 |
+
desc: null
|
| 1062 |
+
value: false
|
| 1063 |
+
memory_snapshot_path:
|
| 1064 |
+
desc: null
|
| 1065 |
+
value: snapshot.pickle
|
| 1066 |
+
record_shapes:
|
| 1067 |
+
desc: null
|
| 1068 |
+
value: false
|
| 1069 |
+
nvtx_ranges:
|
| 1070 |
+
desc: null
|
| 1071 |
+
value: false
|
| 1072 |
+
micro_batch_size:
|
| 1073 |
+
desc: null
|
| 1074 |
+
value: 1
|
| 1075 |
+
global_batch_size:
|
| 1076 |
+
desc: null
|
| 1077 |
+
value: 512
|
| 1078 |
+
rampup_batch_size:
|
| 1079 |
+
desc: null
|
| 1080 |
+
value: null
|
| 1081 |
+
step_batch_size_schedule:
|
| 1082 |
+
desc: null
|
| 1083 |
+
value: null
|
| 1084 |
+
decrease_batch_size_if_needed:
|
| 1085 |
+
desc: null
|
| 1086 |
+
value: false
|
| 1087 |
+
empty_unused_memory_level:
|
| 1088 |
+
desc: null
|
| 1089 |
+
value: 0
|
| 1090 |
+
check_weight_hash_across_dp_replicas_interval:
|
| 1091 |
+
desc: null
|
| 1092 |
+
value: null
|
| 1093 |
+
gpu_sniff_test_interval:
|
| 1094 |
+
desc: null
|
| 1095 |
+
value: null
|
| 1096 |
+
train_sync_interval:
|
| 1097 |
+
desc: null
|
| 1098 |
+
value: null
|
| 1099 |
+
train_iters:
|
| 1100 |
+
desc: null
|
| 1101 |
+
value: 1430512
|
| 1102 |
+
train_samples:
|
| 1103 |
+
desc: null
|
| 1104 |
+
value: null
|
| 1105 |
+
exit_interval:
|
| 1106 |
+
desc: null
|
| 1107 |
+
value: null
|
| 1108 |
+
exit_duration_in_mins:
|
| 1109 |
+
desc: null
|
| 1110 |
+
value: null
|
| 1111 |
+
exit_signal_handler:
|
| 1112 |
+
desc: null
|
| 1113 |
+
value: false
|
| 1114 |
+
exit_signal:
|
| 1115 |
+
desc: null
|
| 1116 |
+
value: SIGTERM
|
| 1117 |
+
exit_signal_handler_for_dataloader:
|
| 1118 |
+
desc: null
|
| 1119 |
+
value: false
|
| 1120 |
+
exit_signal_handler_for_training:
|
| 1121 |
+
desc: null
|
| 1122 |
+
value: false
|
| 1123 |
+
manual_gc:
|
| 1124 |
+
desc: null
|
| 1125 |
+
value: true
|
| 1126 |
+
manual_gc_interval:
|
| 1127 |
+
desc: null
|
| 1128 |
+
value: 10
|
| 1129 |
+
manual_gc_eval:
|
| 1130 |
+
desc: null
|
| 1131 |
+
value: true
|
| 1132 |
+
iterations_to_skip:
|
| 1133 |
+
desc: null
|
| 1134 |
+
value: []
|
| 1135 |
+
check_for_nan_in_loss_and_grad:
|
| 1136 |
+
desc: null
|
| 1137 |
+
value: true
|
| 1138 |
+
check_for_large_grads:
|
| 1139 |
+
desc: null
|
| 1140 |
+
value: false
|
| 1141 |
+
result_rejected_tracker_filename:
|
| 1142 |
+
desc: null
|
| 1143 |
+
value: null
|
| 1144 |
+
tp_comm_overlap_cfg:
|
| 1145 |
+
desc: null
|
| 1146 |
+
value: null
|
| 1147 |
+
masked_softmax_fusion:
|
| 1148 |
+
desc: null
|
| 1149 |
+
value: true
|
| 1150 |
+
bias_gelu_fusion:
|
| 1151 |
+
desc: null
|
| 1152 |
+
value: false
|
| 1153 |
+
bias_swiglu_fusion:
|
| 1154 |
+
desc: null
|
| 1155 |
+
value: true
|
| 1156 |
+
bias_dropout_fusion:
|
| 1157 |
+
desc: null
|
| 1158 |
+
value: true
|
| 1159 |
+
apply_rope_fusion:
|
| 1160 |
+
desc: null
|
| 1161 |
+
value: true
|
| 1162 |
+
rope_type:
|
| 1163 |
+
desc: null
|
| 1164 |
+
value: null
|
| 1165 |
+
use_flash_attn:
|
| 1166 |
+
desc: null
|
| 1167 |
+
value: false
|
| 1168 |
+
optimizer:
|
| 1169 |
+
desc: null
|
| 1170 |
+
value: muon
|
| 1171 |
+
optimizer_cpu_offload:
|
| 1172 |
+
desc: null
|
| 1173 |
+
value: false
|
| 1174 |
+
optimizer_cuda_graph:
|
| 1175 |
+
desc: null
|
| 1176 |
+
value: false
|
| 1177 |
+
optimizer_offload_fraction:
|
| 1178 |
+
desc: null
|
| 1179 |
+
value: 1.0
|
| 1180 |
+
use_torch_optimizer_for_cpu_offload:
|
| 1181 |
+
desc: null
|
| 1182 |
+
value: false
|
| 1183 |
+
overlap_cpu_optimizer_d2h_h2d:
|
| 1184 |
+
desc: null
|
| 1185 |
+
value: false
|
| 1186 |
+
dump_param_to_param_group_map:
|
| 1187 |
+
desc: null
|
| 1188 |
+
value: null
|
| 1189 |
+
pin_cpu_grads:
|
| 1190 |
+
desc: null
|
| 1191 |
+
value: true
|
| 1192 |
+
pin_cpu_params:
|
| 1193 |
+
desc: null
|
| 1194 |
+
value: true
|
| 1195 |
+
dataloader_type:
|
| 1196 |
+
desc: null
|
| 1197 |
+
value: single
|
| 1198 |
+
no_persist_layer_norm:
|
| 1199 |
+
desc: null
|
| 1200 |
+
value: false
|
| 1201 |
+
gradient_accumulation_fusion:
|
| 1202 |
+
desc: null
|
| 1203 |
+
value: true
|
| 1204 |
+
deprecated_use_mcore_models:
|
| 1205 |
+
desc: null
|
| 1206 |
+
value: true
|
| 1207 |
+
perform_rl_step:
|
| 1208 |
+
desc: null
|
| 1209 |
+
value: false
|
| 1210 |
+
rl_prompts_per_eval:
|
| 1211 |
+
desc: null
|
| 1212 |
+
value: 32
|
| 1213 |
+
grpo_prompts_per_step:
|
| 1214 |
+
desc: null
|
| 1215 |
+
value: 32
|
| 1216 |
+
grpo_group_size:
|
| 1217 |
+
desc: null
|
| 1218 |
+
value: 2
|
| 1219 |
+
rl_num_parallel_generations:
|
| 1220 |
+
desc: null
|
| 1221 |
+
value: null
|
| 1222 |
+
rl_num_parallel_generation_batches:
|
| 1223 |
+
desc: null
|
| 1224 |
+
value: null
|
| 1225 |
+
rl_generation_batch_size:
|
| 1226 |
+
desc: null
|
| 1227 |
+
value: null
|
| 1228 |
+
grpo_iterations:
|
| 1229 |
+
desc: null
|
| 1230 |
+
value: 2
|
| 1231 |
+
grpo_clamp_eps_lower:
|
| 1232 |
+
desc: null
|
| 1233 |
+
value: 0.01
|
| 1234 |
+
grpo_clamp_eps_upper:
|
| 1235 |
+
desc: null
|
| 1236 |
+
value: 0.01
|
| 1237 |
+
grpo_kl_beta:
|
| 1238 |
+
desc: null
|
| 1239 |
+
value: 0.001
|
| 1240 |
+
grpo_entropy_term_weight:
|
| 1241 |
+
desc: null
|
| 1242 |
+
value: 0.0
|
| 1243 |
+
grpo_filter_groups_with_same_reward:
|
| 1244 |
+
desc: null
|
| 1245 |
+
value: false
|
| 1246 |
+
langrl_env_config:
|
| 1247 |
+
desc: null
|
| 1248 |
+
value: null
|
| 1249 |
+
rl_default_temperature:
|
| 1250 |
+
desc: null
|
| 1251 |
+
value: 1.0
|
| 1252 |
+
rl_default_top_p:
|
| 1253 |
+
desc: null
|
| 1254 |
+
value: 0
|
| 1255 |
+
rl_default_top_k:
|
| 1256 |
+
desc: null
|
| 1257 |
+
value: -1
|
| 1258 |
+
rl_offload_optimizer_during_inference:
|
| 1259 |
+
desc: null
|
| 1260 |
+
value: false
|
| 1261 |
+
rl_kv_cache_management_mode:
|
| 1262 |
+
desc: null
|
| 1263 |
+
value: persist
|
| 1264 |
+
rl_persist_cuda_graphs:
|
| 1265 |
+
desc: null
|
| 1266 |
+
value: false
|
| 1267 |
+
rl_partial_rollouts:
|
| 1268 |
+
desc: null
|
| 1269 |
+
value: false
|
| 1270 |
+
rl_inference_logprobs_is_correction:
|
| 1271 |
+
desc: null
|
| 1272 |
+
value: false
|
| 1273 |
+
rl_importance_sampling_truncation_coef:
|
| 1274 |
+
desc: null
|
| 1275 |
+
value: null
|
| 1276 |
+
rl_use_sequence_packing:
|
| 1277 |
+
desc: null
|
| 1278 |
+
value: false
|
| 1279 |
+
rl_sequence_packing_max_sequences_per_bin:
|
| 1280 |
+
desc: null
|
| 1281 |
+
value: 50
|
| 1282 |
+
rl_sequence_packing_algo:
|
| 1283 |
+
desc: null
|
| 1284 |
+
value: fifo
|
| 1285 |
+
rl_training_cuda_graphs:
|
| 1286 |
+
desc: null
|
| 1287 |
+
value: false
|
| 1288 |
+
rl_inference_tensor_model_parallel_size:
|
| 1289 |
+
desc: null
|
| 1290 |
+
value: null
|
| 1291 |
+
rl_inference_pipeline_model_parallel_size:
|
| 1292 |
+
desc: null
|
| 1293 |
+
value: null
|
| 1294 |
+
rl_inference_expert_model_parallel_size:
|
| 1295 |
+
desc: null
|
| 1296 |
+
value: null
|
| 1297 |
+
rl_inference_expert_tensor_model_parallel_size:
|
| 1298 |
+
desc: null
|
| 1299 |
+
value: null
|
| 1300 |
+
rl_inference_model_unified_memory_level:
|
| 1301 |
+
desc: null
|
| 1302 |
+
value: 0
|
| 1303 |
+
rl_offload_inference_model_weights_when_idle:
|
| 1304 |
+
desc: null
|
| 1305 |
+
value: false
|
| 1306 |
+
refit_method:
|
| 1307 |
+
desc: null
|
| 1308 |
+
value: gloo
|
| 1309 |
+
rl_verify_model_weights_swap:
|
| 1310 |
+
desc: null
|
| 1311 |
+
value: false
|
| 1312 |
+
rl_parallel_generation_tasks:
|
| 1313 |
+
desc: null
|
| 1314 |
+
value: null
|
| 1315 |
+
rl_skip_bos_token:
|
| 1316 |
+
desc: null
|
| 1317 |
+
value: false
|
| 1318 |
+
rl_inference_parsers:
|
| 1319 |
+
desc: null
|
| 1320 |
+
value: []
|
| 1321 |
+
seed:
|
| 1322 |
+
desc: null
|
| 1323 |
+
value: 1234
|
| 1324 |
+
te_rng_tracker:
|
| 1325 |
+
desc: null
|
| 1326 |
+
value: false
|
| 1327 |
+
inference_rng_tracker:
|
| 1328 |
+
desc: null
|
| 1329 |
+
value: false
|
| 1330 |
+
data_parallel_random_init:
|
| 1331 |
+
desc: null
|
| 1332 |
+
value: false
|
| 1333 |
+
init_method_xavier_uniform:
|
| 1334 |
+
desc: null
|
| 1335 |
+
value: false
|
| 1336 |
+
lr_decay_style:
|
| 1337 |
+
desc: null
|
| 1338 |
+
value: cosine
|
| 1339 |
+
lr_wsd_decay_style:
|
| 1340 |
+
desc: null
|
| 1341 |
+
value: exponential
|
| 1342 |
+
lr_decay_iters:
|
| 1343 |
+
desc: null
|
| 1344 |
+
value: 1430512
|
| 1345 |
+
lr_decay_samples:
|
| 1346 |
+
desc: null
|
| 1347 |
+
value: null
|
| 1348 |
+
lr_wsd_decay_iters:
|
| 1349 |
+
desc: null
|
| 1350 |
+
value: null
|
| 1351 |
+
lr_wsd_decay_samples:
|
| 1352 |
+
desc: null
|
| 1353 |
+
value: null
|
| 1354 |
+
lr_warmup_fraction:
|
| 1355 |
+
desc: null
|
| 1356 |
+
value: null
|
| 1357 |
+
lr_warmup_iters:
|
| 1358 |
+
desc: null
|
| 1359 |
+
value: 2000
|
| 1360 |
+
lr_warmup_samples:
|
| 1361 |
+
desc: null
|
| 1362 |
+
value: 0
|
| 1363 |
+
lr_warmup_init:
|
| 1364 |
+
desc: null
|
| 1365 |
+
value: 0.0
|
| 1366 |
+
override_opt_param_scheduler:
|
| 1367 |
+
desc: null
|
| 1368 |
+
value: false
|
| 1369 |
+
use_checkpoint_opt_param_scheduler:
|
| 1370 |
+
desc: null
|
| 1371 |
+
value: false
|
| 1372 |
+
start_weight_decay:
|
| 1373 |
+
desc: null
|
| 1374 |
+
value: 0.1
|
| 1375 |
+
end_weight_decay:
|
| 1376 |
+
desc: null
|
| 1377 |
+
value: 0.1
|
| 1378 |
+
weight_decay_incr_style:
|
| 1379 |
+
desc: null
|
| 1380 |
+
value: constant
|
| 1381 |
+
lr:
|
| 1382 |
+
desc: null
|
| 1383 |
+
value: 6.0e-05
|
| 1384 |
+
min_lr:
|
| 1385 |
+
desc: null
|
| 1386 |
+
value: 6.0e-06
|
| 1387 |
+
decoupled_lr:
|
| 1388 |
+
desc: null
|
| 1389 |
+
value: null
|
| 1390 |
+
decoupled_min_lr:
|
| 1391 |
+
desc: null
|
| 1392 |
+
value: null
|
| 1393 |
+
save:
|
| 1394 |
+
desc: null
|
| 1395 |
+
value: /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/checkpoints
|
| 1396 |
+
save_interval:
|
| 1397 |
+
desc: null
|
| 1398 |
+
value: 10000
|
| 1399 |
+
save_params_interval:
|
| 1400 |
+
desc: null
|
| 1401 |
+
value: null
|
| 1402 |
+
save_activations_interval:
|
| 1403 |
+
desc: null
|
| 1404 |
+
value: null
|
| 1405 |
+
save_tokens_per_expert_interval:
|
| 1406 |
+
desc: null
|
| 1407 |
+
value: null
|
| 1408 |
+
save_wgrads_interval:
|
| 1409 |
+
desc: null
|
| 1410 |
+
value: null
|
| 1411 |
+
save_dgrads_interval:
|
| 1412 |
+
desc: null
|
| 1413 |
+
value: null
|
| 1414 |
+
save_retain_interval:
|
| 1415 |
+
desc: null
|
| 1416 |
+
value: null
|
| 1417 |
+
load:
|
| 1418 |
+
desc: null
|
| 1419 |
+
value: null
|
| 1420 |
+
load_main_params_from_ckpt:
|
| 1421 |
+
desc: null
|
| 1422 |
+
value: false
|
| 1423 |
+
non_persistent_save_interval:
|
| 1424 |
+
desc: null
|
| 1425 |
+
value: null
|
| 1426 |
+
non_persistent_ckpt_type:
|
| 1427 |
+
desc: null
|
| 1428 |
+
value: null
|
| 1429 |
+
non_persistent_global_ckpt_dir:
|
| 1430 |
+
desc: null
|
| 1431 |
+
value: null
|
| 1432 |
+
non_persistent_local_ckpt_dir:
|
| 1433 |
+
desc: null
|
| 1434 |
+
value: null
|
| 1435 |
+
non_persistent_local_ckpt_algo:
|
| 1436 |
+
desc: null
|
| 1437 |
+
value: fully_parallel
|
| 1438 |
+
finetune:
|
| 1439 |
+
desc: null
|
| 1440 |
+
value: false
|
| 1441 |
+
pretrained_checkpoint:
|
| 1442 |
+
desc: null
|
| 1443 |
+
value: null
|
| 1444 |
+
ckpt_step:
|
| 1445 |
+
desc: null
|
| 1446 |
+
value: null
|
| 1447 |
+
use_checkpoint_args:
|
| 1448 |
+
desc: null
|
| 1449 |
+
value: false
|
| 1450 |
+
use_mp_args_from_checkpoint_args:
|
| 1451 |
+
desc: null
|
| 1452 |
+
value: false
|
| 1453 |
+
use_tokenizer_model_from_checkpoint_args:
|
| 1454 |
+
desc: null
|
| 1455 |
+
value: true
|
| 1456 |
+
exit_on_missing_checkpoint:
|
| 1457 |
+
desc: null
|
| 1458 |
+
value: false
|
| 1459 |
+
ckpt_format:
|
| 1460 |
+
desc: null
|
| 1461 |
+
value: torch_dist
|
| 1462 |
+
auto_detect_ckpt_format:
|
| 1463 |
+
desc: null
|
| 1464 |
+
value: false
|
| 1465 |
+
ckpt_convert_format:
|
| 1466 |
+
desc: null
|
| 1467 |
+
value: null
|
| 1468 |
+
ckpt_convert_save:
|
| 1469 |
+
desc: null
|
| 1470 |
+
value: null
|
| 1471 |
+
ckpt_convert_update_legacy_dist_opt_format:
|
| 1472 |
+
desc: null
|
| 1473 |
+
value: false
|
| 1474 |
+
ckpt_fully_parallel_save:
|
| 1475 |
+
desc: null
|
| 1476 |
+
value: true
|
| 1477 |
+
async_save:
|
| 1478 |
+
desc: null
|
| 1479 |
+
value: false
|
| 1480 |
+
async_strategy:
|
| 1481 |
+
desc: null
|
| 1482 |
+
value: mcore
|
| 1483 |
+
use_persistent_ckpt_worker:
|
| 1484 |
+
desc: null
|
| 1485 |
+
value: false
|
| 1486 |
+
async_ckpt_cpu_priority:
|
| 1487 |
+
desc: null
|
| 1488 |
+
value: 10
|
| 1489 |
+
async_ckpt_io_priority:
|
| 1490 |
+
desc: null
|
| 1491 |
+
value: 3
|
| 1492 |
+
async_ckpt_use_cpu_shm:
|
| 1493 |
+
desc: null
|
| 1494 |
+
value: false
|
| 1495 |
+
ckpt_fully_parallel_load:
|
| 1496 |
+
desc: null
|
| 1497 |
+
value: false
|
| 1498 |
+
ckpt_fully_parallel_load_exchange_algo:
|
| 1499 |
+
desc: null
|
| 1500 |
+
value: broadcast
|
| 1501 |
+
ckpt_fully_parallel_save_process_group:
|
| 1502 |
+
desc: null
|
| 1503 |
+
value: dp
|
| 1504 |
+
ckpt_fully_parallel_load_process_group:
|
| 1505 |
+
desc: null
|
| 1506 |
+
value: dp
|
| 1507 |
+
ckpt_assume_constant_structure:
|
| 1508 |
+
desc: null
|
| 1509 |
+
value: false
|
| 1510 |
+
ckpt_load_validate_sharding_integrity:
|
| 1511 |
+
desc: null
|
| 1512 |
+
value: true
|
| 1513 |
+
strict_fsdp_dtensor_load:
|
| 1514 |
+
desc: null
|
| 1515 |
+
value: true
|
| 1516 |
+
dist_ckpt_strictness:
|
| 1517 |
+
desc: null
|
| 1518 |
+
value: assume_ok_unexpected
|
| 1519 |
+
dist_ckpt_save_pre_mcore_014:
|
| 1520 |
+
desc: null
|
| 1521 |
+
value: false
|
| 1522 |
+
dist_ckpt_optim_fully_reshardable:
|
| 1523 |
+
desc: null
|
| 1524 |
+
value: false
|
| 1525 |
+
distrib_optim_fully_reshardable_mem_efficient:
|
| 1526 |
+
desc: null
|
| 1527 |
+
value: false
|
| 1528 |
+
replication:
|
| 1529 |
+
desc: null
|
| 1530 |
+
value: false
|
| 1531 |
+
replication_jump:
|
| 1532 |
+
desc: null
|
| 1533 |
+
value: null
|
| 1534 |
+
replication_factor:
|
| 1535 |
+
desc: null
|
| 1536 |
+
value: 2
|
| 1537 |
+
verify_integrity:
|
| 1538 |
+
desc: null
|
| 1539 |
+
value: false
|
| 1540 |
+
no_save_optim:
|
| 1541 |
+
desc: null
|
| 1542 |
+
value: null
|
| 1543 |
+
no_save_rng:
|
| 1544 |
+
desc: null
|
| 1545 |
+
value: null
|
| 1546 |
+
no_load_optim:
|
| 1547 |
+
desc: null
|
| 1548 |
+
value: null
|
| 1549 |
+
no_load_rng:
|
| 1550 |
+
desc: null
|
| 1551 |
+
value: null
|
| 1552 |
+
use_dist_ckpt_deprecated:
|
| 1553 |
+
desc: null
|
| 1554 |
+
value: false
|
| 1555 |
+
dist_ckpt_format_deprecated:
|
| 1556 |
+
desc: null
|
| 1557 |
+
value: null
|
| 1558 |
+
dist_ckpt_workers:
|
| 1559 |
+
desc: null
|
| 1560 |
+
value: 1
|
| 1561 |
+
ckpt_fully_parallel_save_deprecated:
|
| 1562 |
+
desc: null
|
| 1563 |
+
value: false
|
| 1564 |
+
grad_reduce_in_bf16:
|
| 1565 |
+
desc: null
|
| 1566 |
+
value: false
|
| 1567 |
+
loss_scale:
|
| 1568 |
+
desc: null
|
| 1569 |
+
value: null
|
| 1570 |
+
initial_loss_scale:
|
| 1571 |
+
desc: null
|
| 1572 |
+
value: 4294967296
|
| 1573 |
+
min_loss_scale:
|
| 1574 |
+
desc: null
|
| 1575 |
+
value: 1.0
|
| 1576 |
+
loss_scale_window:
|
| 1577 |
+
desc: null
|
| 1578 |
+
value: 1000
|
| 1579 |
+
hysteresis:
|
| 1580 |
+
desc: null
|
| 1581 |
+
value: 2
|
| 1582 |
+
attention_softmax_in_fp32:
|
| 1583 |
+
desc: null
|
| 1584 |
+
value: false
|
| 1585 |
+
accumulate_allreduce_grads_in_fp32:
|
| 1586 |
+
desc: null
|
| 1587 |
+
value: true
|
| 1588 |
+
fp16_lm_cross_entropy:
|
| 1589 |
+
desc: null
|
| 1590 |
+
value: false
|
| 1591 |
+
reuse_grad_buf_for_mxfp8_param_ag:
|
| 1592 |
+
desc: null
|
| 1593 |
+
value: false
|
| 1594 |
+
distributed_backend:
|
| 1595 |
+
desc: null
|
| 1596 |
+
value: nccl
|
| 1597 |
+
distributed_timeout_minutes:
|
| 1598 |
+
desc: null
|
| 1599 |
+
value: 60
|
| 1600 |
+
align_grad_reduce:
|
| 1601 |
+
desc: null
|
| 1602 |
+
value: true
|
| 1603 |
+
local_rank:
|
| 1604 |
+
desc: null
|
| 1605 |
+
value: 7
|
| 1606 |
+
lazy_mpu_init:
|
| 1607 |
+
desc: null
|
| 1608 |
+
value: false
|
| 1609 |
+
use_megatron_fsdp:
|
| 1610 |
+
desc: null
|
| 1611 |
+
value: false
|
| 1612 |
+
use_torch_fsdp2:
|
| 1613 |
+
desc: null
|
| 1614 |
+
value: false
|
| 1615 |
+
nccl_communicator_config_path:
|
| 1616 |
+
desc: null
|
| 1617 |
+
value: null
|
| 1618 |
+
use_tp_pp_dp_mapping:
|
| 1619 |
+
desc: null
|
| 1620 |
+
value: false
|
| 1621 |
+
use_gloo_process_groups:
|
| 1622 |
+
desc: null
|
| 1623 |
+
value: true
|
| 1624 |
+
use_sharp:
|
| 1625 |
+
desc: null
|
| 1626 |
+
value: false
|
| 1627 |
+
sharp_enabled_group:
|
| 1628 |
+
desc: null
|
| 1629 |
+
value: null
|
| 1630 |
+
high_priority_stream_groups:
|
| 1631 |
+
desc: null
|
| 1632 |
+
value: []
|
| 1633 |
+
distributed_timeout_seconds_after_init:
|
| 1634 |
+
desc: null
|
| 1635 |
+
value: null
|
| 1636 |
+
flight_recorder_dump_path:
|
| 1637 |
+
desc: null
|
| 1638 |
+
value: null
|
| 1639 |
+
flight_recorder_trace_buffer_size:
|
| 1640 |
+
desc: null
|
| 1641 |
+
value: 2000
|
| 1642 |
+
flight_recorder_dump_on_timeout:
|
| 1643 |
+
desc: null
|
| 1644 |
+
value: true
|
| 1645 |
+
flight_recorder_include_stack_trace:
|
| 1646 |
+
desc: null
|
| 1647 |
+
value: false
|
| 1648 |
+
flight_recorder_include_only_active:
|
| 1649 |
+
desc: null
|
| 1650 |
+
value: true
|
| 1651 |
+
flight_recorder_extra_dump_on_exec:
|
| 1652 |
+
desc: null
|
| 1653 |
+
value: true
|
| 1654 |
+
disable_jit_fuser:
|
| 1655 |
+
desc: null
|
| 1656 |
+
value: false
|
| 1657 |
+
decoder_first_pipeline_num_layers:
|
| 1658 |
+
desc: null
|
| 1659 |
+
value: null
|
| 1660 |
+
decoder_last_pipeline_num_layers:
|
| 1661 |
+
desc: null
|
| 1662 |
+
value: null
|
| 1663 |
+
pipeline_model_parallel_layout:
|
| 1664 |
+
desc: null
|
| 1665 |
+
value: null
|
| 1666 |
+
num_layers_per_virtual_pipeline_stage:
|
| 1667 |
+
desc: null
|
| 1668 |
+
value: null
|
| 1669 |
+
num_virtual_stages_per_pipeline_rank:
|
| 1670 |
+
desc: null
|
| 1671 |
+
value: null
|
| 1672 |
+
overlap_p2p_comm:
|
| 1673 |
+
desc: null
|
| 1674 |
+
value: false
|
| 1675 |
+
overlap_grad_reduce:
|
| 1676 |
+
desc: null
|
| 1677 |
+
value: true
|
| 1678 |
+
ddp_num_buckets:
|
| 1679 |
+
desc: null
|
| 1680 |
+
value: null
|
| 1681 |
+
ddp_bucket_size:
|
| 1682 |
+
desc: null
|
| 1683 |
+
value: 480000000
|
| 1684 |
+
ddp_pad_buckets_for_high_nccl_busbw:
|
| 1685 |
+
desc: null
|
| 1686 |
+
value: true
|
| 1687 |
+
ddp_reduce_scatter_with_fp32_accumulation:
|
| 1688 |
+
desc: null
|
| 1689 |
+
value: false
|
| 1690 |
+
ddp_param_name_patterns_for_fp32_local_accumulation:
|
| 1691 |
+
desc: null
|
| 1692 |
+
value: []
|
| 1693 |
+
ddp_average_in_collective:
|
| 1694 |
+
desc: null
|
| 1695 |
+
value: false
|
| 1696 |
+
overlap_param_gather:
|
| 1697 |
+
desc: null
|
| 1698 |
+
value: false
|
| 1699 |
+
overlap_param_gather_with_optimizer_step:
|
| 1700 |
+
desc: null
|
| 1701 |
+
value: false
|
| 1702 |
+
align_param_gather:
|
| 1703 |
+
desc: null
|
| 1704 |
+
value: false
|
| 1705 |
+
use_distributed_optimizer:
|
| 1706 |
+
desc: null
|
| 1707 |
+
value: false
|
| 1708 |
+
nccl_ub:
|
| 1709 |
+
desc: null
|
| 1710 |
+
value: false
|
| 1711 |
+
disable_symmetric_registration:
|
| 1712 |
+
desc: null
|
| 1713 |
+
value: false
|
| 1714 |
+
fsdp_manual_registration:
|
| 1715 |
+
desc: null
|
| 1716 |
+
value: false
|
| 1717 |
+
create_all_gather_group:
|
| 1718 |
+
desc: null
|
| 1719 |
+
value: false
|
| 1720 |
+
data_parallel_sharding_strategy:
|
| 1721 |
+
desc: null
|
| 1722 |
+
value: optim_grads_params
|
| 1723 |
+
outer_dp_sharding_strategy:
|
| 1724 |
+
desc: null
|
| 1725 |
+
value: no_shard
|
| 1726 |
+
gradient_reduce_div_fusion:
|
| 1727 |
+
desc: null
|
| 1728 |
+
value: true
|
| 1729 |
+
fsdp_double_buffer:
|
| 1730 |
+
desc: null
|
| 1731 |
+
value: false
|
| 1732 |
+
suggested_communication_unit_size:
|
| 1733 |
+
desc: null
|
| 1734 |
+
value: null
|
| 1735 |
+
keep_fp8_transpose_cache:
|
| 1736 |
+
desc: null
|
| 1737 |
+
value: false
|
| 1738 |
+
enable_full_sharding_in_hsdp:
|
| 1739 |
+
desc: null
|
| 1740 |
+
value: false
|
| 1741 |
+
num_distributed_optimizer_instances:
|
| 1742 |
+
desc: null
|
| 1743 |
+
value: 1
|
| 1744 |
+
torch_fsdp2_reshard_after_forward:
|
| 1745 |
+
desc: null
|
| 1746 |
+
value: true
|
| 1747 |
+
cp_comm_type:
|
| 1748 |
+
desc: null
|
| 1749 |
+
value:
|
| 1750 |
+
- p2p
|
| 1751 |
+
fake_process_group:
|
| 1752 |
+
desc: null
|
| 1753 |
+
value: false
|
| 1754 |
+
eval_iters:
|
| 1755 |
+
desc: null
|
| 1756 |
+
value: 0
|
| 1757 |
+
eval_interval:
|
| 1758 |
+
desc: null
|
| 1759 |
+
value: 1000
|
| 1760 |
+
start_eval_at_iter:
|
| 1761 |
+
desc: null
|
| 1762 |
+
value: null
|
| 1763 |
+
eval_global_batch_size:
|
| 1764 |
+
desc: null
|
| 1765 |
+
value: 512
|
| 1766 |
+
eval_micro_batch_size:
|
| 1767 |
+
desc: null
|
| 1768 |
+
value: 1
|
| 1769 |
+
skip_train:
|
| 1770 |
+
desc: null
|
| 1771 |
+
value: false
|
| 1772 |
+
test_mode:
|
| 1773 |
+
desc: null
|
| 1774 |
+
value: false
|
| 1775 |
+
full_validation:
|
| 1776 |
+
desc: null
|
| 1777 |
+
value: false
|
| 1778 |
+
multiple_validation_sets:
|
| 1779 |
+
desc: null
|
| 1780 |
+
value: false
|
| 1781 |
+
validation_set_names:
|
| 1782 |
+
desc: null
|
| 1783 |
+
value: null
|
| 1784 |
+
data_path:
|
| 1785 |
+
desc: null
|
| 1786 |
+
value: null
|
| 1787 |
+
phase_transition_iterations:
|
| 1788 |
+
desc: null
|
| 1789 |
+
value: null
|
| 1790 |
+
split:
|
| 1791 |
+
desc: null
|
| 1792 |
+
value: 100,0,0
|
| 1793 |
+
train_data_path:
|
| 1794 |
+
desc: null
|
| 1795 |
+
value: null
|
| 1796 |
+
valid_data_path:
|
| 1797 |
+
desc: null
|
| 1798 |
+
value: null
|
| 1799 |
+
test_data_path:
|
| 1800 |
+
desc: null
|
| 1801 |
+
value: null
|
| 1802 |
+
data_args_path:
|
| 1803 |
+
desc: null
|
| 1804 |
+
value: /mnt/shared-storage-gpfs2/plm-gpfs/public_data_model/data/dolma3_mix-6T/megatron_olmo3_8192/data_args_path.txt
|
| 1805 |
+
per_split_data_args_path:
|
| 1806 |
+
desc: null
|
| 1807 |
+
value: null
|
| 1808 |
+
per_dataset_sequences_path:
|
| 1809 |
+
desc: null
|
| 1810 |
+
value: /mnt/shared-storage-gpfs2/plm-gpfs/public_data_model/data/dolma3_mix-6T/megatron_olmo3_8192/per_dataset_sequences.json
|
| 1811 |
+
dataloader_fast_cache_load:
|
| 1812 |
+
desc: null
|
| 1813 |
+
value: false
|
| 1814 |
+
dataloader_defer_npy_index_mmap:
|
| 1815 |
+
desc: null
|
| 1816 |
+
value: false
|
| 1817 |
+
data_cache_path:
|
| 1818 |
+
desc: null
|
| 1819 |
+
value: /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/run/c22vq-dddiag256-full-0611113636/data_cache/real
|
| 1820 |
+
mmap_bin_files:
|
| 1821 |
+
desc: null
|
| 1822 |
+
value: false
|
| 1823 |
+
mock_data:
|
| 1824 |
+
desc: null
|
| 1825 |
+
value: false
|
| 1826 |
+
seq_length:
|
| 1827 |
+
desc: null
|
| 1828 |
+
value: 8192
|
| 1829 |
+
encoder_seq_length:
|
| 1830 |
+
desc: null
|
| 1831 |
+
value: 8192
|
| 1832 |
+
decoder_seq_length:
|
| 1833 |
+
desc: null
|
| 1834 |
+
value: null
|
| 1835 |
+
sample_rate:
|
| 1836 |
+
desc: null
|
| 1837 |
+
value: 1.0
|
| 1838 |
+
mask_prob:
|
| 1839 |
+
desc: null
|
| 1840 |
+
value: 0.15
|
| 1841 |
+
short_seq_prob:
|
| 1842 |
+
desc: null
|
| 1843 |
+
value: 0.1
|
| 1844 |
+
num_workers:
|
| 1845 |
+
desc: null
|
| 1846 |
+
value: 2
|
| 1847 |
+
reset_position_ids:
|
| 1848 |
+
desc: null
|
| 1849 |
+
value: false
|
| 1850 |
+
reset_attention_mask:
|
| 1851 |
+
desc: null
|
| 1852 |
+
value: false
|
| 1853 |
+
eod_mask_loss:
|
| 1854 |
+
desc: null
|
| 1855 |
+
value: false
|
| 1856 |
+
create_attention_mask_in_dataloader:
|
| 1857 |
+
desc: null
|
| 1858 |
+
value: false
|
| 1859 |
+
num_dataset_builder_threads:
|
| 1860 |
+
desc: null
|
| 1861 |
+
value: 1
|
| 1862 |
+
object_storage_cache_path:
|
| 1863 |
+
desc: null
|
| 1864 |
+
value: null
|
| 1865 |
+
mid_level_dataset_surplus:
|
| 1866 |
+
desc: null
|
| 1867 |
+
value: 0.005
|
| 1868 |
+
allow_ambiguous_pad_tokens:
|
| 1869 |
+
desc: null
|
| 1870 |
+
value: false
|
| 1871 |
+
fim_data:
|
| 1872 |
+
desc: null
|
| 1873 |
+
value: false
|
| 1874 |
+
fim_rate:
|
| 1875 |
+
desc: null
|
| 1876 |
+
value: 0.5
|
| 1877 |
+
fim_spm_rate:
|
| 1878 |
+
desc: null
|
| 1879 |
+
value: 0.5
|
| 1880 |
+
fim_split_sample:
|
| 1881 |
+
desc: null
|
| 1882 |
+
value: null
|
| 1883 |
+
fim_fragment_rate:
|
| 1884 |
+
desc: null
|
| 1885 |
+
value: null
|
| 1886 |
+
fim_no_prefix:
|
| 1887 |
+
desc: null
|
| 1888 |
+
value: null
|
| 1889 |
+
fim_prefix_token:
|
| 1890 |
+
desc: null
|
| 1891 |
+
value: <fim_prefix>
|
| 1892 |
+
fim_middle_token:
|
| 1893 |
+
desc: null
|
| 1894 |
+
value: <fim_middle>
|
| 1895 |
+
fim_suffix_token:
|
| 1896 |
+
desc: null
|
| 1897 |
+
value: <fim_suffix>
|
| 1898 |
+
fim_pad_token:
|
| 1899 |
+
desc: null
|
| 1900 |
+
value: <fim_pad>
|
| 1901 |
+
fim_eod_token:
|
| 1902 |
+
desc: null
|
| 1903 |
+
value: <|endoftext|>
|
| 1904 |
+
vocab_size:
|
| 1905 |
+
desc: null
|
| 1906 |
+
value: 100278
|
| 1907 |
+
padded_vocab_size:
|
| 1908 |
+
desc: null
|
| 1909 |
+
value: 100278
|
| 1910 |
+
vocab_file:
|
| 1911 |
+
desc: null
|
| 1912 |
+
value: null
|
| 1913 |
+
merge_file:
|
| 1914 |
+
desc: null
|
| 1915 |
+
value: null
|
| 1916 |
+
vocab_extra_ids:
|
| 1917 |
+
desc: null
|
| 1918 |
+
value: 0
|
| 1919 |
+
tokenizer_type:
|
| 1920 |
+
desc: null
|
| 1921 |
+
value: HuggingFaceTokenizer
|
| 1922 |
+
tokenizer_model:
|
| 1923 |
+
desc: null
|
| 1924 |
+
value: /mnt/shared-storage-gpfs2/plm-gpfs/public_data_model/model/Olmo-3-1025-7B-stage1-step1413814
|
| 1925 |
+
metadata_path:
|
| 1926 |
+
desc: null
|
| 1927 |
+
value: null
|
| 1928 |
+
special_tokens:
|
| 1929 |
+
desc: null
|
| 1930 |
+
value: null
|
| 1931 |
+
tiktoken_pattern:
|
| 1932 |
+
desc: null
|
| 1933 |
+
value: null
|
| 1934 |
+
tiktoken_num_special_tokens:
|
| 1935 |
+
desc: null
|
| 1936 |
+
value: 1000
|
| 1937 |
+
tokenizer_sentencepiece_legacy:
|
| 1938 |
+
desc: null
|
| 1939 |
+
value: false
|
| 1940 |
+
tokenizer_hf_no_use_fast:
|
| 1941 |
+
desc: null
|
| 1942 |
+
value: false
|
| 1943 |
+
tokenizer_hf_no_include_special_tokens:
|
| 1944 |
+
desc: null
|
| 1945 |
+
value: false
|
| 1946 |
+
trust_remote_code:
|
| 1947 |
+
desc: null
|
| 1948 |
+
value: false
|
| 1949 |
+
null_tokenizer_eod_id:
|
| 1950 |
+
desc: null
|
| 1951 |
+
value: null
|
| 1952 |
+
null_tokenizer_pad_id:
|
| 1953 |
+
desc: null
|
| 1954 |
+
value: -1
|
| 1955 |
+
chat_template:
|
| 1956 |
+
desc: null
|
| 1957 |
+
value: null
|
| 1958 |
+
adlr_autoresume:
|
| 1959 |
+
desc: null
|
| 1960 |
+
value: false
|
| 1961 |
+
adlr_autoresume_interval:
|
| 1962 |
+
desc: null
|
| 1963 |
+
value: 1000
|
| 1964 |
+
ict_head_size:
|
| 1965 |
+
desc: null
|
| 1966 |
+
value: null
|
| 1967 |
+
biencoder_projection_dim:
|
| 1968 |
+
desc: null
|
| 1969 |
+
value: 0
|
| 1970 |
+
biencoder_shared_query_context_model:
|
| 1971 |
+
desc: null
|
| 1972 |
+
value: false
|
| 1973 |
+
ict_load:
|
| 1974 |
+
desc: null
|
| 1975 |
+
value: null
|
| 1976 |
+
bert_load:
|
| 1977 |
+
desc: null
|
| 1978 |
+
value: null
|
| 1979 |
+
titles_data_path:
|
| 1980 |
+
desc: null
|
| 1981 |
+
value: null
|
| 1982 |
+
query_in_block_prob:
|
| 1983 |
+
desc: null
|
| 1984 |
+
value: 0.1
|
| 1985 |
+
use_one_sent_docs:
|
| 1986 |
+
desc: null
|
| 1987 |
+
value: false
|
| 1988 |
+
evidence_data_path:
|
| 1989 |
+
desc: null
|
| 1990 |
+
value: null
|
| 1991 |
+
retriever_report_topk_accuracies:
|
| 1992 |
+
desc: null
|
| 1993 |
+
value: []
|
| 1994 |
+
retriever_score_scaling:
|
| 1995 |
+
desc: null
|
| 1996 |
+
value: false
|
| 1997 |
+
block_data_path:
|
| 1998 |
+
desc: null
|
| 1999 |
+
value: null
|
| 2000 |
+
embedding_path:
|
| 2001 |
+
desc: null
|
| 2002 |
+
value: null
|
| 2003 |
+
indexer_batch_size:
|
| 2004 |
+
desc: null
|
| 2005 |
+
value: 128
|
| 2006 |
+
indexer_log_interval:
|
| 2007 |
+
desc: null
|
| 2008 |
+
value: 1000
|
| 2009 |
+
num_classes:
|
| 2010 |
+
desc: null
|
| 2011 |
+
value: 1000
|
| 2012 |
+
img_h:
|
| 2013 |
+
desc: null
|
| 2014 |
+
value: 224
|
| 2015 |
+
img_w:
|
| 2016 |
+
desc: null
|
| 2017 |
+
value: 224
|
| 2018 |
+
num_channels:
|
| 2019 |
+
desc: null
|
| 2020 |
+
value: 3
|
| 2021 |
+
patch_dim:
|
| 2022 |
+
desc: null
|
| 2023 |
+
value: 16
|
| 2024 |
+
classes_fraction:
|
| 2025 |
+
desc: null
|
| 2026 |
+
value: 1.0
|
| 2027 |
+
data_per_class_fraction:
|
| 2028 |
+
desc: null
|
| 2029 |
+
value: 1.0
|
| 2030 |
+
data_sharding:
|
| 2031 |
+
desc: null
|
| 2032 |
+
value: true
|
| 2033 |
+
head_lr_mult:
|
| 2034 |
+
desc: null
|
| 2035 |
+
value: 1.0
|
| 2036 |
+
vision_pretraining:
|
| 2037 |
+
desc: null
|
| 2038 |
+
value: false
|
| 2039 |
+
vision_pretraining_type:
|
| 2040 |
+
desc: null
|
| 2041 |
+
value: classify
|
| 2042 |
+
vision_backbone_type:
|
| 2043 |
+
desc: null
|
| 2044 |
+
value: vit
|
| 2045 |
+
swin_backbone_type:
|
| 2046 |
+
desc: null
|
| 2047 |
+
value: tiny
|
| 2048 |
+
mask_type:
|
| 2049 |
+
desc: null
|
| 2050 |
+
value: random
|
| 2051 |
+
mask_factor:
|
| 2052 |
+
desc: null
|
| 2053 |
+
value: 1.0
|
| 2054 |
+
iter_per_epoch:
|
| 2055 |
+
desc: null
|
| 2056 |
+
value: 1250
|
| 2057 |
+
dino_local_img_size:
|
| 2058 |
+
desc: null
|
| 2059 |
+
value: 96
|
| 2060 |
+
dino_local_crops_number:
|
| 2061 |
+
desc: null
|
| 2062 |
+
value: 10
|
| 2063 |
+
dino_head_hidden_size:
|
| 2064 |
+
desc: null
|
| 2065 |
+
value: 2048
|
| 2066 |
+
dino_bottleneck_size:
|
| 2067 |
+
desc: null
|
| 2068 |
+
value: 256
|
| 2069 |
+
dino_freeze_last_layer:
|
| 2070 |
+
desc: null
|
| 2071 |
+
value: 1
|
| 2072 |
+
dino_norm_last_layer:
|
| 2073 |
+
desc: null
|
| 2074 |
+
value: false
|
| 2075 |
+
dino_warmup_teacher_temp:
|
| 2076 |
+
desc: null
|
| 2077 |
+
value: 0.04
|
| 2078 |
+
dino_teacher_temp:
|
| 2079 |
+
desc: null
|
| 2080 |
+
value: 0.07
|
| 2081 |
+
dino_warmup_teacher_temp_epochs:
|
| 2082 |
+
desc: null
|
| 2083 |
+
value: 30
|
| 2084 |
+
num_experts:
|
| 2085 |
+
desc: null
|
| 2086 |
+
value: null
|
| 2087 |
+
moe_layer_freq:
|
| 2088 |
+
desc: null
|
| 2089 |
+
value: 1
|
| 2090 |
+
moe_use_upcycling:
|
| 2091 |
+
desc: null
|
| 2092 |
+
value: false
|
| 2093 |
+
moe_router_load_balancing_type:
|
| 2094 |
+
desc: null
|
| 2095 |
+
value: aux_loss
|
| 2096 |
+
moe_aux_loss_coeff:
|
| 2097 |
+
desc: null
|
| 2098 |
+
value: 0.0
|
| 2099 |
+
moe_upcycling_granularity:
|
| 2100 |
+
desc: null
|
| 2101 |
+
value: 1
|
| 2102 |
+
q_lora_rank:
|
| 2103 |
+
desc: null
|
| 2104 |
+
value: null
|
| 2105 |
+
kv_lora_rank:
|
| 2106 |
+
desc: null
|
| 2107 |
+
value: 32
|
| 2108 |
+
qk_head_dim:
|
| 2109 |
+
desc: null
|
| 2110 |
+
value: 128
|
| 2111 |
+
qk_pos_emb_head_dim:
|
| 2112 |
+
desc: null
|
| 2113 |
+
value: 64
|
| 2114 |
+
v_head_dim:
|
| 2115 |
+
desc: null
|
| 2116 |
+
value: 128
|
| 2117 |
+
rotary_scaling_factor:
|
| 2118 |
+
desc: null
|
| 2119 |
+
value: 1.0
|
| 2120 |
+
mscale:
|
| 2121 |
+
desc: null
|
| 2122 |
+
value: 1.0
|
| 2123 |
+
mscale_all_dim:
|
| 2124 |
+
desc: null
|
| 2125 |
+
value: 0.0
|
| 2126 |
+
cache_mla_latents:
|
| 2127 |
+
desc: null
|
| 2128 |
+
value: false
|
| 2129 |
+
mla_down_proj_fusion:
|
| 2130 |
+
desc: null
|
| 2131 |
+
value: false
|
| 2132 |
+
linear_attention_freq:
|
| 2133 |
+
desc: null
|
| 2134 |
+
value: null
|
| 2135 |
+
heterogeneous_layers_config_path:
|
| 2136 |
+
desc: null
|
| 2137 |
+
value: null
|
| 2138 |
+
heterogeneous_layers_config_encoded_json:
|
| 2139 |
+
desc: null
|
| 2140 |
+
value: null
|
| 2141 |
+
log_interval:
|
| 2142 |
+
desc: null
|
| 2143 |
+
value: 1
|
| 2144 |
+
log_params_norm:
|
| 2145 |
+
desc: null
|
| 2146 |
+
value: false
|
| 2147 |
+
log_throughput:
|
| 2148 |
+
desc: null
|
| 2149 |
+
value: true
|
| 2150 |
+
log_progress:
|
| 2151 |
+
desc: null
|
| 2152 |
+
value: false
|
| 2153 |
+
timing_log_level:
|
| 2154 |
+
desc: null
|
| 2155 |
+
value: 0
|
| 2156 |
+
timing_log_option:
|
| 2157 |
+
desc: null
|
| 2158 |
+
value: minmax
|
| 2159 |
+
tensorboard_dir:
|
| 2160 |
+
desc: null
|
| 2161 |
+
value: /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/run/c22vq-dddiag256-full-0611113636/tensorboard
|
| 2162 |
+
tensorboard_log_interval:
|
| 2163 |
+
desc: null
|
| 2164 |
+
value: 1
|
| 2165 |
+
tensorboard_queue_size:
|
| 2166 |
+
desc: null
|
| 2167 |
+
value: 1000
|
| 2168 |
+
log_timers_to_tensorboard:
|
| 2169 |
+
desc: null
|
| 2170 |
+
value: true
|
| 2171 |
+
log_loss_scale_to_tensorboard:
|
| 2172 |
+
desc: null
|
| 2173 |
+
value: true
|
| 2174 |
+
log_validation_ppl_to_tensorboard:
|
| 2175 |
+
desc: null
|
| 2176 |
+
value: false
|
| 2177 |
+
log_memory_to_tensorboard:
|
| 2178 |
+
desc: null
|
| 2179 |
+
value: false
|
| 2180 |
+
log_memory_interval:
|
| 2181 |
+
desc: null
|
| 2182 |
+
value: null
|
| 2183 |
+
log_device_memory_used:
|
| 2184 |
+
desc: null
|
| 2185 |
+
value: false
|
| 2186 |
+
log_num_zeros_in_grad:
|
| 2187 |
+
desc: null
|
| 2188 |
+
value: false
|
| 2189 |
+
log_max_attention_logit:
|
| 2190 |
+
desc: null
|
| 2191 |
+
value: false
|
| 2192 |
+
barrier_with_L1_time:
|
| 2193 |
+
desc: null
|
| 2194 |
+
value: true
|
| 2195 |
+
log_world_size_to_tensorboard:
|
| 2196 |
+
desc: null
|
| 2197 |
+
value: false
|
| 2198 |
+
wandb_project:
|
| 2199 |
+
desc: null
|
| 2200 |
+
value: ConceptLM
|
| 2201 |
+
wandb_exp_name:
|
| 2202 |
+
desc: null
|
| 2203 |
+
value: c22vq-dddiag256-full-0611113636
|
| 2204 |
+
wandb_save_dir:
|
| 2205 |
+
desc: null
|
| 2206 |
+
value: /mnt/shared-storage-gpfs2/plm-gpfs/ylliu/scale_up/train_conceptlm_V22_olmo_7B/full_train_dd_diag_256g_0611/runs/dddiag256_full_1f85baf_0611113636/run/c22vq-dddiag256-full-0611113636/wandb
|
| 2207 |
+
wandb_entity:
|
| 2208 |
+
desc: null
|
| 2209 |
+
value: iammi-nanjing-university
|
| 2210 |
+
logging_level:
|
| 2211 |
+
desc: null
|
| 2212 |
+
value: null
|
| 2213 |
+
log_energy:
|
| 2214 |
+
desc: null
|
| 2215 |
+
value: false
|
| 2216 |
+
log_straggler:
|
| 2217 |
+
desc: null
|
| 2218 |
+
value: false
|
| 2219 |
+
straggler_ctrlr_port:
|
| 2220 |
+
desc: null
|
| 2221 |
+
value: 65535
|
| 2222 |
+
straggler_minmax_count:
|
| 2223 |
+
desc: null
|
| 2224 |
+
value: 1
|
| 2225 |
+
disable_straggler_on_startup:
|
| 2226 |
+
desc: null
|
| 2227 |
+
value: false
|
| 2228 |
+
run_workload_inspector_server:
|
| 2229 |
+
desc: null
|
| 2230 |
+
value: false
|
| 2231 |
+
inference_batch_times_seqlen_threshold:
|
| 2232 |
+
desc: null
|
| 2233 |
+
value: -1
|
| 2234 |
+
max_tokens_to_oom:
|
| 2235 |
+
desc: null
|
| 2236 |
+
value: 12000
|
| 2237 |
+
output_bert_embeddings:
|
| 2238 |
+
desc: null
|
| 2239 |
+
value: false
|
| 2240 |
+
bert_embedder_type:
|
| 2241 |
+
desc: null
|
| 2242 |
+
value: megatron
|
| 2243 |
+
cuda_graph_scope:
|
| 2244 |
+
desc: null
|
| 2245 |
+
value: []
|
| 2246 |
+
use_legacy_static_engine:
|
| 2247 |
+
desc: null
|
| 2248 |
+
value: false
|
| 2249 |
+
inference_max_requests:
|
| 2250 |
+
desc: null
|
| 2251 |
+
value: 8
|
| 2252 |
+
inference_max_seq_length:
|
| 2253 |
+
desc: null
|
| 2254 |
+
value: 2560
|
| 2255 |
+
inference_dynamic_batching:
|
| 2256 |
+
desc: null
|
| 2257 |
+
value: false
|
| 2258 |
+
inference_dynamic_batching_buffer_size_gb:
|
| 2259 |
+
desc: null
|
| 2260 |
+
value: 40.0
|
| 2261 |
+
inference_dynamic_batching_paused_buffer_size_gb:
|
| 2262 |
+
desc: null
|
| 2263 |
+
value: null
|
| 2264 |
+
inference_dynamic_batching_mamba_memory_ratio:
|
| 2265 |
+
desc: null
|
| 2266 |
+
value: null
|
| 2267 |
+
inference_dynamic_batching_block_size:
|
| 2268 |
+
desc: null
|
| 2269 |
+
value: 256
|
| 2270 |
+
inference_dynamic_batching_max_requests:
|
| 2271 |
+
desc: null
|
| 2272 |
+
value: null
|
| 2273 |
+
inference_dynamic_batching_max_tokens:
|
| 2274 |
+
desc: null
|
| 2275 |
+
value: null
|
| 2276 |
+
inference_dynamic_batching_num_cuda_graphs:
|
| 2277 |
+
desc: null
|
| 2278 |
+
value: 16
|
| 2279 |
+
inference_dynamic_batching_track_paused_request_events:
|
| 2280 |
+
desc: null
|
| 2281 |
+
value: false
|
| 2282 |
+
inference_dynamic_batching_track_generated_token_events:
|
| 2283 |
+
desc: null
|
| 2284 |
+
value: false
|
| 2285 |
+
decode_only_cuda_graphs:
|
| 2286 |
+
desc: null
|
| 2287 |
+
value: false
|
| 2288 |
+
inference_dynamic_batching_unified_memory_level:
|
| 2289 |
+
desc: null
|
| 2290 |
+
value: 0
|
| 2291 |
+
enable_chunked_prefill:
|
| 2292 |
+
desc: null
|
| 2293 |
+
value: false
|
| 2294 |
+
num_speculative_tokens:
|
| 2295 |
+
desc: null
|
| 2296 |
+
value: 0
|
| 2297 |
+
inference_dynamic_batching_enable_prefix_caching:
|
| 2298 |
+
desc: null
|
| 2299 |
+
value: false
|
| 2300 |
+
inference_dynamic_batching_prefix_caching_eviction_policy:
|
| 2301 |
+
desc: null
|
| 2302 |
+
value: ref_zero
|
| 2303 |
+
inference_dynamic_batching_prefix_caching_coordinator_policy:
|
| 2304 |
+
desc: null
|
| 2305 |
+
value: first_prefix_block
|
| 2306 |
+
inference_dynamic_batching_prefix_caching_routing_alpha:
|
| 2307 |
+
desc: null
|
| 2308 |
+
value: 0.5
|
| 2309 |
+
inference_dynamic_batching_prefix_caching_mamba_gb:
|
| 2310 |
+
desc: null
|
| 2311 |
+
value: null
|
| 2312 |
+
inference_dynamic_batching_cuda_graph_max_tokens:
|
| 2313 |
+
desc: null
|
| 2314 |
+
value: 16384
|
| 2315 |
+
inference_dynamic_batching_cuda_graph_mixed_prefill_count:
|
| 2316 |
+
desc: null
|
| 2317 |
+
value: 16
|
| 2318 |
+
inference_dynamic_batching_sampling_backend:
|
| 2319 |
+
desc: null
|
| 2320 |
+
value: torch
|
| 2321 |
+
inference_logging_step_interval:
|
| 2322 |
+
desc: null
|
| 2323 |
+
value: 0
|
| 2324 |
+
inference_text_gen_server_logging:
|
| 2325 |
+
desc: null
|
| 2326 |
+
value: false
|
| 2327 |
+
inference_wandb_logging:
|
| 2328 |
+
desc: null
|
| 2329 |
+
value: false
|
| 2330 |
+
inference_coordinator_port:
|
| 2331 |
+
desc: null
|
| 2332 |
+
value: null
|
| 2333 |
+
mamba_inference_conv_states_dtype:
|
| 2334 |
+
desc: null
|
| 2335 |
+
value: torch.bfloat16
|
| 2336 |
+
mamba_inference_ssm_states_dtype:
|
| 2337 |
+
desc: null
|
| 2338 |
+
value: torch.bfloat16
|
| 2339 |
+
inference_use_synchronous_zmq_collectives:
|
| 2340 |
+
desc: null
|
| 2341 |
+
value: false
|
| 2342 |
+
inference_disable_ep_consensus:
|
| 2343 |
+
desc: null
|
| 2344 |
+
value: false
|
| 2345 |
+
fp4_param_gather:
|
| 2346 |
+
desc: null
|
| 2347 |
+
value: false
|
| 2348 |
+
fp8_param_gather:
|
| 2349 |
+
desc: null
|
| 2350 |
+
value: false
|
| 2351 |
+
te_precision_config_file:
|
| 2352 |
+
desc: null
|
| 2353 |
+
value: null
|
| 2354 |
+
enable_experimental:
|
| 2355 |
+
desc: null
|
| 2356 |
+
value: false
|
| 2357 |
+
spec:
|
| 2358 |
+
desc: null
|
| 2359 |
+
value:
|
| 2360 |
+
- megatron.core.models.gpt.olmo3_layer_specs
|
| 2361 |
+
- olmo3_layer_spec
|
| 2362 |
+
hybrid_layer_pattern:
|
| 2363 |
+
desc: null
|
| 2364 |
+
value: null
|
| 2365 |
+
hybrid_override_pattern:
|
| 2366 |
+
desc: null
|
| 2367 |
+
value: null
|
| 2368 |
+
yaml_cfg:
|
| 2369 |
+
desc: null
|
| 2370 |
+
value: null
|
| 2371 |
+
use_precision_aware_optimizer:
|
| 2372 |
+
desc: null
|
| 2373 |
+
value: false
|
| 2374 |
+
main_grads_dtype:
|
| 2375 |
+
desc: null
|
| 2376 |
+
value: torch.float32
|
| 2377 |
+
main_params_dtype:
|
| 2378 |
+
desc: null
|
| 2379 |
+
value: torch.float32
|
| 2380 |
+
exp_avg_dtype:
|
| 2381 |
+
desc: null
|
| 2382 |
+
value: torch.float32
|
| 2383 |
+
exp_avg_sq_dtype:
|
| 2384 |
+
desc: null
|
| 2385 |
+
value: torch.float32
|
| 2386 |
+
megatron_fsdp_main_params_dtype:
|
| 2387 |
+
desc: null
|
| 2388 |
+
value: torch.float32
|
| 2389 |
+
megatron_fsdp_main_grads_dtype:
|
| 2390 |
+
desc: null
|
| 2391 |
+
value: null
|
| 2392 |
+
megatron_fsdp_grad_comm_dtype:
|
| 2393 |
+
desc: null
|
| 2394 |
+
value: null
|
| 2395 |
+
enable_one_logger:
|
| 2396 |
+
desc: null
|
| 2397 |
+
value: true
|
| 2398 |
+
one_logger_project:
|
| 2399 |
+
desc: null
|
| 2400 |
+
value: megatron-lm
|
| 2401 |
+
one_logger_run_name:
|
| 2402 |
+
desc: null
|
| 2403 |
+
value: null
|
| 2404 |
+
one_logger_async:
|
| 2405 |
+
desc: null
|
| 2406 |
+
value: false
|
| 2407 |
+
app_tag_run_name:
|
| 2408 |
+
desc: null
|
| 2409 |
+
value: null
|
| 2410 |
+
app_tag_run_version:
|
| 2411 |
+
desc: null
|
| 2412 |
+
value: 0.0.0
|
| 2413 |
+
inprocess_restart:
|
| 2414 |
+
desc: null
|
| 2415 |
+
value: false
|
| 2416 |
+
inprocess_max_iterations:
|
| 2417 |
+
desc: null
|
| 2418 |
+
value: null
|
| 2419 |
+
inprocess_monitor_thread_interval:
|
| 2420 |
+
desc: null
|
| 2421 |
+
value: 1.0
|
| 2422 |
+
inprocess_monitor_process_interval:
|
| 2423 |
+
desc: null
|
| 2424 |
+
value: 1.0
|
| 2425 |
+
inprocess_progress_watchdog_interval:
|
| 2426 |
+
desc: null
|
| 2427 |
+
value: 1.0
|
| 2428 |
+
inprocess_heartbeat_interval:
|
| 2429 |
+
desc: null
|
| 2430 |
+
value: 30
|
| 2431 |
+
inprocess_soft_timeout:
|
| 2432 |
+
desc: null
|
| 2433 |
+
value: 60
|
| 2434 |
+
inprocess_hard_timeout:
|
| 2435 |
+
desc: null
|
| 2436 |
+
value: 90
|
| 2437 |
+
inprocess_heartbeat_timeout:
|
| 2438 |
+
desc: null
|
| 2439 |
+
value: 60
|
| 2440 |
+
inprocess_barrier_timeout:
|
| 2441 |
+
desc: null
|
| 2442 |
+
value: 120
|
| 2443 |
+
inprocess_completion_timeout:
|
| 2444 |
+
desc: null
|
| 2445 |
+
value: 120
|
| 2446 |
+
inprocess_last_call_wait:
|
| 2447 |
+
desc: null
|
| 2448 |
+
value: 1
|
| 2449 |
+
inprocess_termination_grace_time:
|
| 2450 |
+
desc: null
|
| 2451 |
+
value: 1
|
| 2452 |
+
inprocess_granularity:
|
| 2453 |
+
desc: null
|
| 2454 |
+
value: node
|
| 2455 |
+
inprocess_active_world_size:
|
| 2456 |
+
desc: null
|
| 2457 |
+
value: 256
|
| 2458 |
+
inprocess_empty_cuda_cache:
|
| 2459 |
+
desc: null
|
| 2460 |
+
value: false
|
| 2461 |
+
enable_ft_package:
|
| 2462 |
+
desc: null
|
| 2463 |
+
value: false
|
| 2464 |
+
calc_ft_timeouts:
|
| 2465 |
+
desc: null
|
| 2466 |
+
value: false
|
| 2467 |
+
ft_num_warmup_iters:
|
| 2468 |
+
desc: null
|
| 2469 |
+
value: 5
|
| 2470 |
+
error_injection_rate:
|
| 2471 |
+
desc: null
|
| 2472 |
+
value: 0
|
| 2473 |
+
error_injection_type:
|
| 2474 |
+
desc: null
|
| 2475 |
+
value: transient_error
|
| 2476 |
+
rerun_mode:
|
| 2477 |
+
desc: null
|
| 2478 |
+
value: disabled
|
| 2479 |
+
check_for_spiky_loss:
|
| 2480 |
+
desc: null
|
| 2481 |
+
value: false
|
| 2482 |
+
enable_msc:
|
| 2483 |
+
desc: null
|
| 2484 |
+
value: true
|
| 2485 |
+
sft:
|
| 2486 |
+
desc: null
|
| 2487 |
+
value: false
|
| 2488 |
+
sft_tokenizer_prompt_format:
|
| 2489 |
+
desc: null
|
| 2490 |
+
value: nemotron-h-aligned
|
| 2491 |
+
fault_injector_ranks:
|
| 2492 |
+
desc: null
|
| 2493 |
+
value: null
|
| 2494 |
+
fault_injector_num_ranks:
|
| 2495 |
+
desc: null
|
| 2496 |
+
value: null
|
| 2497 |
+
fault_injector_fault_types:
|
| 2498 |
+
desc: null
|
| 2499 |
+
value: null
|
| 2500 |
+
fault_injector_fault_probabilities:
|
| 2501 |
+
desc: null
|
| 2502 |
+
value: null
|
| 2503 |
+
fault_injector_fault_delay:
|
| 2504 |
+
desc: null
|
| 2505 |
+
value: null
|
| 2506 |
+
fault_injector_delay_start_iteration:
|
| 2507 |
+
desc: null
|
| 2508 |
+
value: null
|
| 2509 |
+
fault_injector_mtti_seconds:
|
| 2510 |
+
desc: null
|
| 2511 |
+
value: null
|
| 2512 |
+
fault_injector_offset_seconds:
|
| 2513 |
+
desc: null
|
| 2514 |
+
value: null
|
| 2515 |
+
fault_injector_seed:
|
| 2516 |
+
desc: null
|
| 2517 |
+
value: null
|
| 2518 |
+
export_model_type:
|
| 2519 |
+
desc: null
|
| 2520 |
+
value: GPTModel
|
| 2521 |
+
export_legacy_megatron:
|
| 2522 |
+
desc: null
|
| 2523 |
+
value: false
|
| 2524 |
+
export_te_mcore_model:
|
| 2525 |
+
desc: null
|
| 2526 |
+
value: false
|
| 2527 |
+
export_default_te_spec:
|
| 2528 |
+
desc: null
|
| 2529 |
+
value: false
|
| 2530 |
+
export_force_local_attention:
|
| 2531 |
+
desc: null
|
| 2532 |
+
value: false
|
| 2533 |
+
export_kv_cache_quant:
|
| 2534 |
+
desc: null
|
| 2535 |
+
value: none
|
| 2536 |
+
export_real_quant_cfg:
|
| 2537 |
+
desc: null
|
| 2538 |
+
value: None
|
| 2539 |
+
export_quant_cfg:
|
| 2540 |
+
desc: null
|
| 2541 |
+
value: null
|
| 2542 |
+
export_kd_teacher_load:
|
| 2543 |
+
desc: null
|
| 2544 |
+
value: null
|
| 2545 |
+
export_kd_teacher_model_config:
|
| 2546 |
+
desc: null
|
| 2547 |
+
value: null
|
| 2548 |
+
export_kd_teacher_ckpt_format:
|
| 2549 |
+
desc: null
|
| 2550 |
+
value: null
|
| 2551 |
+
export_kd_cfg:
|
| 2552 |
+
desc: null
|
| 2553 |
+
value: null
|
| 2554 |
+
finetune_hf_dataset:
|
| 2555 |
+
desc: null
|
| 2556 |
+
value: null
|
| 2557 |
+
finetune_data_split:
|
| 2558 |
+
desc: null
|
| 2559 |
+
value: train
|
| 2560 |
+
export_qk_l2_norm:
|
| 2561 |
+
desc: null
|
| 2562 |
+
value: false
|
| 2563 |
+
export_moe_apply_probs_on_input:
|
| 2564 |
+
desc: null
|
| 2565 |
+
value: false
|
| 2566 |
+
export_offline_model:
|
| 2567 |
+
desc: null
|
| 2568 |
+
value: false
|
| 2569 |
+
modelopt_enabled:
|
| 2570 |
+
desc: null
|
| 2571 |
+
value: false
|
| 2572 |
+
enable_gpt_oss:
|
| 2573 |
+
desc: null
|
| 2574 |
+
value: false
|
| 2575 |
+
conceptlm_backbone:
|
| 2576 |
+
desc: null
|
| 2577 |
+
value: olmo3
|
| 2578 |
+
conceptlm_encoder_layers:
|
| 2579 |
+
desc: null
|
| 2580 |
+
value: 16
|
| 2581 |
+
conceptlm_special_layers:
|
| 2582 |
+
desc: null
|
| 2583 |
+
value: 8
|
| 2584 |
+
conceptlm_decoder_layers:
|
| 2585 |
+
desc: null
|
| 2586 |
+
value: 16
|
| 2587 |
+
conceptlm_chunk_size:
|
| 2588 |
+
desc: null
|
| 2589 |
+
value: 4
|
| 2590 |
+
conceptlm_shift_feature:
|
| 2591 |
+
desc: null
|
| 2592 |
+
value: true
|
| 2593 |
+
conceptlm_chunk_merge_method:
|
| 2594 |
+
desc: null
|
| 2595 |
+
value: meanpooling
|
| 2596 |
+
conceptlm_enable_chunk_dualpath_smoothing:
|
| 2597 |
+
desc: null
|
| 2598 |
+
value: false
|
| 2599 |
+
conceptlm_chunk_dualpath_alpha_init:
|
| 2600 |
+
desc: null
|
| 2601 |
+
value: 0.5
|
| 2602 |
+
conceptlm_layer_norm_option:
|
| 2603 |
+
desc: null
|
| 2604 |
+
value: normed_add
|
| 2605 |
+
conceptlm_fusion_norm_alpha_init:
|
| 2606 |
+
desc: null
|
| 2607 |
+
value: 0.1
|
| 2608 |
+
conceptlm_fusion_alpha_init:
|
| 2609 |
+
desc: null
|
| 2610 |
+
value: 1.0
|
| 2611 |
+
conceptlm_hlm_ffn_hidden_size:
|
| 2612 |
+
desc: null
|
| 2613 |
+
value: null
|
| 2614 |
+
conceptlm_hlm_attention_mode:
|
| 2615 |
+
desc: null
|
| 2616 |
+
value: backbone_window
|
| 2617 |
+
conceptlm_hlm_loss_weight:
|
| 2618 |
+
desc: null
|
| 2619 |
+
value: 1.0
|
| 2620 |
+
conceptlm_vq_loss_weight:
|
| 2621 |
+
desc: null
|
| 2622 |
+
value: 1.0
|
| 2623 |
+
conceptlm_v22_vq_codebook_size:
|
| 2624 |
+
desc: null
|
| 2625 |
+
value: 128
|
| 2626 |
+
conceptlm_v22_vq_num_codebooks:
|
| 2627 |
+
desc: null
|
| 2628 |
+
value: 32
|
| 2629 |
+
conceptlm_v22_vq_commitment_cost:
|
| 2630 |
+
desc: null
|
| 2631 |
+
value: 0.25
|
| 2632 |
+
conceptlm_v22_vq_merge_mode:
|
| 2633 |
+
desc: null
|
| 2634 |
+
value: raw_logits
|
| 2635 |
+
conceptlm_v22_vq_hlm_loss_type:
|
| 2636 |
+
desc: null
|
| 2637 |
+
value: mse
|
| 2638 |
+
conceptlm_v22_vq_detach_hlm_target:
|
| 2639 |
+
desc: null
|
| 2640 |
+
value: true
|
| 2641 |
+
conceptlm_v22_vq_lr_mult:
|
| 2642 |
+
desc: null
|
| 2643 |
+
value: 1.0
|
| 2644 |
+
conceptlm_v22_vq_highlevel_lr_mult:
|
| 2645 |
+
desc: null
|
| 2646 |
+
value: 1.0
|
| 2647 |
+
conceptlm_v21_dd_two_route_add:
|
| 2648 |
+
desc: null
|
| 2649 |
+
value: true
|
| 2650 |
+
conceptlm_v21_dd_two_route_add_concept_source:
|
| 2651 |
+
desc: null
|
| 2652 |
+
value: final
|
| 2653 |
+
conceptlm_v21_dd_two_route_add_enable_raw_concept_route:
|
| 2654 |
+
desc: null
|
| 2655 |
+
value: false
|
| 2656 |
+
conceptlm_v21_dd_two_route_add_enable_final_concept_route:
|
| 2657 |
+
desc: null
|
| 2658 |
+
value: true
|
| 2659 |
+
conceptlm_v21_dd_two_route_add_beta_init:
|
| 2660 |
+
desc: null
|
| 2661 |
+
value: 0.05
|
| 2662 |
+
conceptlm_v21_dd_two_route_add_every_n_layers:
|
| 2663 |
+
desc: null
|
| 2664 |
+
value: 1
|
| 2665 |
+
conceptlm_v21_dd_two_route_add_concept_route_first_n:
|
| 2666 |
+
desc: null
|
| 2667 |
+
value: -1
|
| 2668 |
+
conceptlm_v21_dd_two_route_add_decoder_hidden_size:
|
| 2669 |
+
desc: null
|
| 2670 |
+
value: 0
|
| 2671 |
+
conceptlm_v21_dd_two_route_add_concept_hidden_size:
|
| 2672 |
+
desc: null
|
| 2673 |
+
value: 0
|
| 2674 |
+
conceptlm_v21_dd_two_route_add_use_softmax:
|
| 2675 |
+
desc: null
|
| 2676 |
+
value: false
|
| 2677 |
+
conceptlm_v21_dd_two_route_add_disable_decoder_dd:
|
| 2678 |
+
desc: null
|
| 2679 |
+
value: false
|
| 2680 |
+
conceptlm_v21_dd_two_route_add_decoder_use_layernorm:
|
| 2681 |
+
desc: null
|
| 2682 |
+
value: false
|
| 2683 |
+
conceptlm_v21_dd_two_route_add_decoder_use_softmax:
|
| 2684 |
+
desc: null
|
| 2685 |
+
value: true
|
| 2686 |
+
conceptlm_v21_dd_two_route_add_concept_use_layernorm:
|
| 2687 |
+
desc: null
|
| 2688 |
+
value: true
|
| 2689 |
+
conceptlm_v21_dd_encoder_self_dd:
|
| 2690 |
+
desc: null
|
| 2691 |
+
value: true
|
| 2692 |
+
conceptlm_v21_dd_encoder_self_dd_every_n_layers:
|
| 2693 |
+
desc: null
|
| 2694 |
+
value: 1
|
| 2695 |
+
conceptlm_v21_dd_encoder_self_dd_hidden_size:
|
| 2696 |
+
desc: null
|
| 2697 |
+
value: 0
|
| 2698 |
+
conceptlm_v21_dd_encoder_self_dd_use_layernorm:
|
| 2699 |
+
desc: null
|
| 2700 |
+
value: false
|
| 2701 |
+
conceptlm_v21_dd_concept_self_dd:
|
| 2702 |
+
desc: null
|
| 2703 |
+
value: true
|
| 2704 |
+
conceptlm_v21_dd_concept_self_dd_every_n_layers:
|
| 2705 |
+
desc: null
|
| 2706 |
+
value: 1
|
| 2707 |
+
conceptlm_v21_dd_concept_self_dd_hidden_size:
|
| 2708 |
+
desc: null
|
| 2709 |
+
value: 0
|
| 2710 |
+
conceptlm_v21_dd_concept_self_dd_use_layernorm:
|
| 2711 |
+
desc: null
|
| 2712 |
+
value: false
|
| 2713 |
+
conceptlm_v21_enable_full_residual_flow:
|
| 2714 |
+
desc: null
|
| 2715 |
+
value: false
|
| 2716 |
+
conceptlm_v21_enable_concept_read_encoder:
|
| 2717 |
+
desc: null
|
| 2718 |
+
value: true
|
| 2719 |
+
conceptlm_v21_enable_decoder_read_encoder:
|
| 2720 |
+
desc: null
|
| 2721 |
+
value: true
|
| 2722 |
+
conceptlm_v21_enable_decoder_read_concept:
|
| 2723 |
+
desc: null
|
| 2724 |
+
value: true
|
| 2725 |
+
conceptlm_v21_concept_read_encoder_first_n:
|
| 2726 |
+
desc: null
|
| 2727 |
+
value: -1
|
| 2728 |
+
conceptlm_v21_decoder_read_encoder_first_n:
|
| 2729 |
+
desc: null
|
| 2730 |
+
value: -1
|
| 2731 |
+
conceptlm_v21_residual_flow_beta_init:
|
| 2732 |
+
desc: null
|
| 2733 |
+
value: 0.02
|
| 2734 |
+
conceptlm_v21_residual_flow_route_hidden_size:
|
| 2735 |
+
desc: null
|
| 2736 |
+
value: 0
|
| 2737 |
+
conceptlm_v21_residual_flow_route_use_softmax:
|
| 2738 |
+
desc: null
|
| 2739 |
+
value: true
|
| 2740 |
+
conceptlm_v21_residual_flow_source_use_layernorm:
|
| 2741 |
+
desc: null
|
| 2742 |
+
value: true
|
| 2743 |
+
conceptlm_v21_residual_flow_shared_source_norm:
|
| 2744 |
+
desc: null
|
| 2745 |
+
value: true
|
| 2746 |
+
conceptlm_v21_final_read_concept_gate:
|
| 2747 |
+
desc: null
|
| 2748 |
+
value: true
|
| 2749 |
+
conceptlm_v21_final_read_concept_gate_init_final:
|
| 2750 |
+
desc: null
|
| 2751 |
+
value: 0.5
|
| 2752 |
+
conceptlm_v21_final_read_concept_gate_target_final:
|
| 2753 |
+
desc: null
|
| 2754 |
+
value: 0.5
|
| 2755 |
+
conceptlm_v21_final_read_concept_gate_reg_weight:
|
| 2756 |
+
desc: null
|
| 2757 |
+
value: 0.0
|
| 2758 |
+
conceptlm_v21_compile_residual_flow_routes:
|
| 2759 |
+
desc: null
|
| 2760 |
+
value: true
|
| 2761 |
+
conceptlm_v21_compile_dd_routes:
|
| 2762 |
+
desc: null
|
| 2763 |
+
value: true
|
| 2764 |
+
_is_global_batch_size_explicitly_specified:
|
| 2765 |
+
desc: null
|
| 2766 |
+
value: true
|
| 2767 |
+
rank:
|
| 2768 |
+
desc: null
|
| 2769 |
+
value: 255
|
| 2770 |
+
world_size:
|
| 2771 |
+
desc: null
|
| 2772 |
+
value: 256
|
| 2773 |
+
use_dist_ckpt:
|
| 2774 |
+
desc: null
|
| 2775 |
+
value: true
|
| 2776 |
+
transformer_pipeline_model_parallel_size:
|
| 2777 |
+
desc: null
|
| 2778 |
+
value: 1
|
| 2779 |
+
data_parallel_size:
|
| 2780 |
+
desc: null
|
| 2781 |
+
value: 256
|
| 2782 |
+
virtual_pipeline_model_parallel_size:
|
| 2783 |
+
desc: null
|
| 2784 |
+
value: null
|
| 2785 |
+
params_dtype:
|
| 2786 |
+
desc: null
|
| 2787 |
+
value: torch.bfloat16
|
| 2788 |
+
consumed_train_samples:
|
| 2789 |
+
desc: null
|
| 2790 |
+
value: 0
|
| 2791 |
+
skipped_train_samples:
|
| 2792 |
+
desc: null
|
| 2793 |
+
value: 0
|
| 2794 |
+
consumed_valid_samples:
|
| 2795 |
+
desc: null
|
| 2796 |
+
value: 0
|
| 2797 |
+
variable_seq_lengths:
|
| 2798 |
+
desc: null
|
| 2799 |
+
value: false
|
| 2800 |
+
use_layer_wise_distributed_optimizer:
|
| 2801 |
+
desc: null
|
| 2802 |
+
value: true
|
| 2803 |
+
slurm_job_name:
|
| 2804 |
+
desc: null
|
| 2805 |
+
value: N/A
|
vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|