Upload train_config.py with huggingface_hub

train_config.py

@@ -0,0 +1,77 @@
+import os
+from deepspec.trainer import Qwen3DSparkTrainer
+BASE_TB_DIR = os.path.expanduser("~/tensorboard")
+BASE_CKPT_DIR = os.path.expanduser("~/checkpoints")
+project_name = "deepspec"
+exp_name = "dspark_block15_qwen36_27b_online"
+seed = 42
+
+model = dict(
+    target_model_name_or_path="/mnt/t5evo/models/Qwen3.6-27B",
+    block_size=15,
+    num_draft_layers=5,
+    # Qwen3.5-0.8B: 24 layers (6 full-attention), hidden 1024
+    target_layer_ids=[1, 16, 31, 46, 61],
+    # free embedding slot: tokenizer len 248077 < embed 248320 (same family trick as ornith_9b)
+    mask_token_id=248200,
+    init_draft_from="/mnt/data/dspark-test/head_q36_27b_dflash_init/model.safetensors",
+    num_anchors=512,
+    markov_rank=256,
+    # match z-lab DFlash warm-start geometry (differs from target-derived defaults)
+    head_dim=128,
+    num_attention_heads=32,
+    num_key_value_heads=8,
+    intermediate_size=17408,
+    markov_head_type="vanilla",
+    confidence_head_alpha=1.0,
+    confidence_head_with_markov=True,
+    loss_decay_gamma=4.0,
+    ce_loss_alpha=0.1,
+    l1_loss_alpha=0.9,
+)
+
+train = dict(
+    trainer_cls=Qwen3DSparkTrainer,
+    lr=5.0e-4,
+    warmup_ratio=0.04,
+    weight_decay=0.0,
+    precision="bf16",
+    local_batch_size=1,
+    global_batch_size=128,   # single Pro 6000; 0.8B target fits with room
+    num_train_epochs=5,
+    max_train_steps=None,
+    max_grad_norm=1.0,
+    sharding_strategy="no_shard",
+    torch_compile=True,
+)
+
+logging = dict(
+    logging_steps=10,
+    checkpointing_steps=200,
+)
+
+data = dict(
+    online_target=True,
+    target_cache_path=None,
+    train_data_paths=[
+        "/mnt/data/DeepSpec/train_datasets/q27_regen_12k.jsonl",
+    ],
+    min_loss_tokens=14,
+    chat_template="qwen",
+    max_length=2048,
+    num_workers=4,
+)
+
+
+def finalize_cfg(cfg):
+    logging_cfg = dict(cfg["logging"])
+    project_name = str(cfg["project_name"])
+    exp_name = str(cfg["exp_name"])
+    logging_cfg["checkpoint_dir"] = os.path.join(BASE_CKPT_DIR, project_name, exp_name)
+    logging_cfg["tensorboard_dir"] = os.path.join(BASE_TB_DIR, project_name, exp_name)
+    cfg["logging"] = logging_cfg
+    return cfg
+
+
+# --opts overrides applied at save time
+train['num_train_epochs'] = 3