#!/usr/bin/env python3 """Debug E4B inference using numpy/safetensors to compare with Swift Metal.""" import numpy as np import json from pathlib import Path from safetensors.numpy import safe_open MODEL_PATH = Path("/Users/accusys/MarkBase12B/models/E4B-MarkBase") def load_config(): with open(MODEL_PATH / "config.json") as f: return json.load(f) def bf16_to_f32(bf16_data): """Convert BF16 uint16 array to float32.""" bf16 = bf16_data.view(np.uint16) f32_bits = bf16.astype(np.uint32) << 16 return f32_bits.view(np.float32) def load_tensor(name): """Load tensor from safetensors, converting BF16 to F32.""" model_file = MODEL_PATH / "model.safetensors" with safe_open(model_file, framework="numpy") as f: data = f.get_tensor(name) # Check if dtype name contains 'bfloat' or is uint16 for scales/biases dtype_str = str(data.dtype) if 'bfloat' in dtype_str.lower() or dtype_str == 'e4m3bfloat16': return bf16_to_f32(data) if data.dtype == np.uint32: return data # Keep uint32 for weights return data.astype(np.float32) def debug_inference(): print("=" * 50) print("NumPy E4B Debug Inference") print("=" * 50) config = load_config() text_config = config["text_config"] print(f"\nConfig: hidden={text_config['hidden_size']}, layers={text_config['num_hidden_layers']}") print(f" head_dim={text_config['head_dim']}, n_heads={text_config['num_attention_heads']}") print(f" n_kv_heads={text_config['num_key_value_heads']}") hidden_size = text_config['hidden_size'] n_heads = text_config['num_attention_heads'] n_kv_heads = text_config['num_key_value_heads'] head_dim = text_config['head_dim'] print("\n--- Loading weights ---") # Embedding embed_weight = load_tensor("language_model.model.embed_tokens.weight") embed_scales = load_tensor("language_model.model.embed_tokens.scales") embed_biases = load_tensor("language_model.model.embed_tokens.biases") print(f"embed_weight: shape={embed_weight.shape}, dtype={embed_weight.dtype}") print(f"embed_scales: shape={embed_scales.shape}, dtype={embed_scales.dtype}") print(f"embed_biases: shape={embed_biases.shape}, dtype={embed_biases.dtype}") # BOS token = 2 bos_token = 2 vocab_size, packed_dim = embed_weight.shape num_groups = hidden_size // 64 print(f"\n--- Dequantizing BOS embedding (token={bos_token}) ---") w_row = embed_weight[bos_token].astype(np.uint32) s_row = embed_scales[bos_token] b_row = embed_biases[bos_token] embedding = np.zeros(hidden_size, dtype=np.float32) for g in range(num_groups): scale = s_row[g] bias = b_row[g] for j in range(64): packed_idx = g * 8 + j // 8 shift = (j % 8) * 4 qval = int((w_row[packed_idx] >> shift) & 0xF) embedding[g * 64 + j] = qval * scale + bias print(f"embedding: shape={embedding.shape}") print(f" min={embedding.min():.6f}, max={embedding.max():.6f}") print(f" mean={embedding.mean():.6f}") print(f" first 20: {embedding[:20].tolist()}") # Embedding scale embed_scale = np.sqrt(float(hidden_size)) embedding_scaled = embedding * embed_scale print(f"\nAfter scaling (scale={embed_scale:.2f}):") print(f" min={embedding_scaled.min():.6f}, max={embedding_scaled.max():.6f}") print(f" first 20: {embedding_scaled[:20].tolist()}") # Layer 0 weights print("\n--- Layer 0 weights ---") prefix = "language_model.model.layers.0" input_norm_w = load_tensor(f"{prefix}.input_layernorm.weight") print(f"input_norm: shape={input_norm_w.shape}, first 5: {input_norm_w[:5].tolist()}") q_proj_w = load_tensor(f"{prefix}.self_attn.q_proj.weight") q_proj_s = load_tensor(f"{prefix}.self_attn.q_proj.scales") q_proj_b = load_tensor(f"{prefix}.self_attn.q_proj.biases") k_proj_w = load_tensor(f"{prefix}.self_attn.k_proj.weight") k_proj_s = load_tensor(f"{prefix}.self_attn.k_proj.scales") k_proj_b = load_tensor(f"{prefix}.self_attn.k_proj.biases") v_proj_w = load_tensor(f"{prefix}.self_attn.v_proj.weight") v_proj_s = load_tensor(f"{prefix}.self_attn.v_proj.scales") v_proj_b = load_tensor(f"{prefix}.self_attn.v_proj.biases") print(f"q_proj: w={q_proj_w.shape}, s={q_proj_s.shape}, b={q_proj_b.shape}") print(f"k_proj: w={k_proj_w.shape}, s={k_proj_s.shape}, b={k_proj_b.shape}") print(f"v_proj: w={v_proj_w.shape}, s={v_proj_s.shape}, b={v_proj_b.shape}") q_norm_w = load_tensor(f"{prefix}.self_attn.q_norm.weight") k_norm_w = load_tensor(f"{prefix}.self_attn.k_norm.weight") print(f"q_norm: shape={q_norm_w.shape}, first 5: {q_norm_w[:5].tolist()}") print(f"k_norm: shape={k_norm_w.shape}, first 5: {k_norm_w[:5].tolist()}") o_proj_w = load_tensor(f"{prefix}.self_attn.o_proj.weight") o_proj_s = load_tensor(f"{prefix}.self_attn.o_proj.scales") o_proj_b = load_tensor(f"{prefix}.self_attn.o_proj.biases") print(f"o_proj: w={o_proj_w.shape}, s={o_proj_s.shape}, b={o_proj_b.shape}") # --- Forward --- print("\n--- Forward pass ---") hidden = embedding_scaled # 1. Input RMS norm eps = 1e-6 ss = np.sum(hidden * hidden) / hidden_size rms = np.sqrt(ss + eps) hidden_normed = hidden / rms * input_norm_w print(f"\nInput norm: rms={rms:.6f}") print(f" min={hidden_normed.min():.6f}, max={hidden_normed.max():.6f}") print(f" first 20: {hidden_normed[:20].tolist()}") # 2. Q, K, V projections def dequantize_matmul(x, w, s, b): out_dim, packed_dim = w.shape in_dim = packed_dim * 8 num_groups = in_dim // 64 result = np.zeros(out_dim, dtype=np.float32) for out_idx in range(out_dim): sum_val = 0.0 for g in range(num_groups): scale = s[out_idx, g] bias_q = b[out_idx, g] for j in range(64): packed_idx = g * 8 + j // 8 shift = (j % 8) * 4 qval = int((w[out_idx, packed_idx] >> shift) & 0xF) dq = qval * scale + bias_q sum_val += dq * x[g * 64 + j] result[out_idx] = sum_val return result print("\nQ projection...") q = dequantize_matmul(hidden_normed, q_proj_w.astype(np.uint32), q_proj_s, q_proj_b) print(f"Q: shape={q.shape}") print(f" min={q.min():.6f}, max={q.max():.6f}") print(f" first 20: {q[:20].tolist()}") print("\nK projection...") k = dequantize_matmul(hidden_normed, k_proj_w.astype(np.uint32), k_proj_s, k_proj_b) print(f"K: shape={k.shape}") print(f" min={k.min():.6f}, max={k.max():.6f}") print(f" first 20: {k[:20].tolist()}") print("\nV projection...") v = dequantize_matmul(hidden_normed, v_proj_w.astype(np.uint32), v_proj_s, v_proj_b) print(f"V: shape={v.shape}") print(f" min={v.min():.6f}, max={v.max():.6f}") print(f" first 20: {v[:20].tolist()}") # 3. Q, K norms print("\nQ norm (per-head)...") q_reshaped = q.reshape(n_heads, head_dim) q_normed = np.zeros_like(q_reshaped) for h in range(n_heads): h_q = q_reshaped[h] ss = np.sum(h_q * h_q) / head_dim rms = np.sqrt(ss + eps) q_normed[h] = h_q / rms * q_norm_w q_normed = q_normed.flatten() print(f"Q normed: shape={q_normed.shape}") print(f" min={q_normed.min():.6f}, max={q_normed.max():.6f}") print(f" first 20: {q_normed[:20].tolist()}") print("\nK norm (per-head)...") k_reshaped = k.reshape(n_kv_heads, head_dim) k_normed = np.zeros_like(k_reshaped) for h in range(n_kv_heads): h_k = k_reshaped[h] ss = np.sum(h_k * h_k) / head_dim rms = np.sqrt(ss + eps) k_normed[h] = h_k / rms * k_norm_w k_normed = k_normed.flatten() print(f"K normed: shape={k_normed.shape}") print(f" min={k_normed.min():.6f}, max={k_normed.max():.6f}") print(f" first 20: {k_normed[:20].tolist()}") # 4. Attention (position 0, only attend to self) print("\n--- Attention at position 0 ---") attn_scale = 1.0 / np.sqrt(float(head_dim)) attn_out = np.zeros(n_heads * head_dim, dtype=np.float32) for h in range(n_heads): kv_head = h % n_kv_heads h_q = q_normed[h * head_dim : (h+1) * head_dim] h_k = k_normed[kv_head * head_dim : (kv_head+1) * head_dim] h_v = v[kv_head * head_dim : (kv_head+1) * head_dim] score = float(np.sum(h_q * h_k) * attn_scale) print(f" Head {h}: score={score:.6f}") # At position 0, softmax(score) = 1.0 for self-attention attn_out[h * head_dim : (h+1) * head_dim] = h_v print(f"\nAttention output: shape={attn_out.shape}") print(f" min={attn_out.min():.6f}, max={attn_out.max():.6f}") print(f" first 20: {attn_out[:20].tolist()}") # 5. O projection print("\nO projection...") o_out = dequantize_matmul(attn_out, o_proj_w.astype(np.uint32), o_proj_s, o_proj_b) print(f"O output: shape={o_out.shape}") print(f" min={o_out.min():.6f}, max={o_out.max():.6f}") print(f" first 20: {o_out[:20].tolist()}") # 6. Residual hidden_after_attn = hidden + o_out print(f"\nAfter attention residual:") print(f" min={hidden_after_attn.min():.6f}, max={hidden_after_attn.max():.6f}") print(f" first 20: {hidden_after_attn[:20].tolist()}") print("\n" + "=" * 50) print("KEY VALUES TO COMPARE WITH SWIFT:") print("=" * 50) print("\n1. Embedding (scaled):") print(f" first 5: {embedding_scaled[:5].tolist()}") print("\n2. After input norm:") print(f" first 5: {hidden_normed[:5].tolist()}") print("\n3. Q after projection:") print(f" first 5: {q[:5].tolist()}") print("\n4. Q after norm:") print(f" first 5: {q_normed[:5].tolist()}") print("\n5. Attention output:") print(f" first 5: {attn_out[:5].tolist()}") print("\n6. O projection output:") print(f" first 5: {o_out[:5].tolist()}") print("\n7. Hidden after attention:") print(f" first 5: {hidden_after_attn[:5].tolist()}") if __name__ == "__main__": debug_inference()