Add V4-Nano config (~523M params for testing)

configs/config_nano.json

@@ -0,0 +1,55 @@
+{
+    "model_type": "deepseek_v4",
+    "architectures": ["DeepSeekV4ForCausalLM"],
+    "_comment": "DeepSeek-V4-Nano: Minimal config for testing and validation (~523M total params, ~200M active)",
+    "vocab_size": 129280,
+    "hidden_size": 1024,
+    "num_hidden_layers": 7,
+    "num_attention_heads": 16,
+    "q_lora_rank": 384,
+    "head_dim": 128,
+    "qk_rope_head_dim": 32,
+    "num_key_value_heads": 1,
+    "o_groups": 4,
+    "o_lora_rank": 256,
+    "window_size": 64,
+    "compress_ratios": [128, 4, 128, 4, 128, 4, 0],
+    "compress_rope_theta": 160000.0,
+    "index_n_heads": 16,
+    "index_head_dim": 64,
+    "index_topk": 128,
+    "moe_intermediate_size": 1024,
+    "n_routed_experts": 8,
+    "n_shared_experts": 1,
+    "num_experts_per_tok": 2,
+    "num_hash_layers": 1,
+    "scoring_func": "sqrtsoftplus",
+    "topk_method": "noaux_tc",
+    "norm_topk_prob": true,
+    "routed_scaling_factor": 2.5,
+    "swiglu_limit": 10.0,
+    "expert_dtype": "bf16",
+    "hc_mult": 4,
+    "hc_sinkhorn_iters": 10,
+    "hc_eps": 1e-6,
+    "num_nextn_predict_layers": 1,
+    "rms_norm_eps": 1e-6,
+    "hidden_act": "silu",
+    "rope_theta": 10000.0,
+    "rope_scaling": {
+        "type": "yarn",
+        "factor": 4,
+        "original_max_position_embeddings": 4096,
+        "beta_fast": 32,
+        "beta_slow": 1
+    },
+    "max_position_embeddings": 16384,
+    "initializer_range": 0.02,
+    "use_cache": true,
+    "tie_word_embeddings": false,
+    "bos_token_id": 0,
+    "eos_token_id": 1,
+    "torch_dtype": "float32",
+    "attention_bias": false,
+    "attention_dropout": 0.0
+}