Add DeepSeek-V4 configuration class

configuration_deepseek_v4.py

@@ -0,0 +1,186 @@
+"""
+DeepSeek-V4 Configuration
+=========================
+HuggingFace-compatible configuration for the DeepSeek-V4 architecture.
+
+Key innovations over DeepSeek-V3:
+- Compressed Sparse Attention (CSA) with sliding window + learned KV compression
+- Manifold-Constrained Hyper-Connections (mHC) replacing residual connections  
+- Hash routing for first N layers + sqrtsoftplus scoring
+- Grouped low-rank output projection
+- Attention sink (learnable per-head bias)
+"""
+
+from transformers import PretrainedConfig
+
+
+class DeepSeekV4Config(PretrainedConfig):
+    """Configuration class for the DeepSeek-V4 model.
+    
+    This config captures ALL architectural parameters from the DeepSeek-V4 technical report
+    and official inference code, enabling faithful reproduction.
+    """
+    
+    model_type = "deepseek_v4"
+    
+    def __init__(
+        self,
+        # Core dimensions
+        vocab_size: int = 129280,
+        hidden_size: int = 7168,
+        num_hidden_layers: int = 61,
+        num_attention_heads: int = 128,
+        
+        # MLA (Multi-head Latent Attention) parameters
+        q_lora_rank: int = 1536,
+        head_dim: int = 512,            # unified KV head dim (V4 unifies nope+rope+v)
+        qk_rope_head_dim: int = 64,     # decoupled RoPE dimension
+        num_key_value_heads: int = 1,    # effectively MQA at compression level
+        
+        # Grouped output projection
+        o_groups: int = 16,              # number of output groups
+        o_lora_rank: int = 1024,         # per-group output compression rank
+        
+        # Compressed Sparse Attention
+        window_size: int = 128,          # sliding window size
+        compress_ratios: tuple = None,   # per-layer compression ratios (0=no compression)
+        compress_rope_theta: float = 160000.0,  # RoPE theta for compressed KV
+        
+        # Indexer (learned sparse attention selection)
+        index_n_heads: int = 64,
+        index_head_dim: int = 128,
+        index_topk: int = 1024,
+        
+        # Attention sink
+        attention_bias: bool = False,    # standard bias in linear layers
+        attention_dropout: float = 0.0,
+        
+        # MoE parameters
+        moe_intermediate_size: int = 3072,
+        n_routed_experts: int = 384,
+        n_shared_experts: int = 1,
+        num_experts_per_tok: int = 6,
+        num_hash_layers: int = 3,        # first N layers use hash routing
+        scoring_func: str = "sqrtsoftplus",  # NEW in V4: sqrt(softplus(x))
+        topk_method: str = "noaux_tc",
+        norm_topk_prob: bool = True,
+        routed_scaling_factor: float = 2.5,
+        swiglu_limit: float = 10.0,      # NEW in V4: clamp SwiGLU activations
+        
+        # Expert quantization
+        expert_dtype: str = "fp4",       # NEW in V4: FP4 expert weights
+        
+        # Hyper-Connections (mHC)
+        hc_mult: int = 4,               # number of hidden state copies
+        hc_sinkhorn_iters: int = 20,     # Sinkhorn normalization iterations
+        hc_eps: float = 1e-6,
+        
+        # Multi-Token Prediction
+        num_nextn_predict_layers: int = 1,
+        
+        # Normalization
+        rms_norm_eps: float = 1e-6,
+        hidden_act: str = "silu",
+        
+        # RoPE parameters
+        rope_theta: float = 10000.0,
+        rope_scaling: dict = None,
+        max_position_embeddings: int = 1048576,  # 1M context
+        
+        # Quantization config (for inference)
+        quantization_config: dict = None,
+        
+        # Standard HF config
+        initializer_range: float = 0.02,
+        use_cache: bool = True,
+        tie_word_embeddings: bool = False,
+        bos_token_id: int = 0,
+        eos_token_id: int = 1,
+        torch_dtype: str = "bfloat16",
+        
+        **kwargs,
+    ):
+        super().__init__(
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+        
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        
+        # MLA
+        self.q_lora_rank = q_lora_rank
+        self.head_dim = head_dim
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.num_key_value_heads = num_key_value_heads
+        
+        # Grouped output
+        self.o_groups = o_groups
+        self.o_lora_rank = o_lora_rank
+        
+        # Compressed Sparse Attention
+        self.window_size = window_size
+        if compress_ratios is not None:
+            self.compress_ratios = list(compress_ratios)
+        else:
+            # Default: V4-Pro pattern — alternating 128 and 4 compression ratios
+            # First 2 layers: 128 (HCA only), then alternating 4,128 pairs, last layer: 0
+            self.compress_ratios = [128, 128] + [4, 128] * 29 + [4, 0]
+        self.compress_rope_theta = compress_rope_theta
+        
+        # Indexer
+        self.index_n_heads = index_n_heads
+        self.index_head_dim = index_head_dim
+        self.index_topk = index_topk
+        
+        # Attention
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        
+        # MoE
+        self.moe_intermediate_size = moe_intermediate_size
+        self.n_routed_experts = n_routed_experts
+        self.n_shared_experts = n_shared_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.num_hash_layers = num_hash_layers
+        self.scoring_func = scoring_func
+        self.topk_method = topk_method
+        self.norm_topk_prob = norm_topk_prob
+        self.routed_scaling_factor = routed_scaling_factor
+        self.swiglu_limit = swiglu_limit
+        self.expert_dtype = expert_dtype
+        
+        # Hyper-Connections
+        self.hc_mult = hc_mult
+        self.hc_sinkhorn_iters = hc_sinkhorn_iters
+        self.hc_eps = hc_eps
+        
+        # MTP
+        self.num_nextn_predict_layers = num_nextn_predict_layers
+        
+        # Normalization
+        self.rms_norm_eps = rms_norm_eps
+        self.hidden_act = hidden_act
+        
+        # RoPE
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling or {
+            "type": "yarn",
+            "factor": 16,
+            "original_max_position_embeddings": 65536,
+            "beta_fast": 32,
+            "beta_slow": 1,
+        }
+        self.max_position_embeddings = max_position_embeddings
+        
+        # Quantization
+        self.quantization_config = quantization_config
+        
+        # Standard
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+        self.torch_dtype = torch_dtype