#!/usr/bin/env python3
"""Compare Swift Metal vs Python NumPy layer 0 outputs at each stage."""

import numpy as np
import json
from pathlib import Path
import struct

MODEL_PATH = Path("/Users/accusys/MarkBase12B/models/E4B-MarkBase")

def load_safetensors_raw():
    """Load safetensors file and return raw data + metadata."""
    model_file = MODEL_PATH / "model.safetensors"
    with open(model_file, 'rb') as f:
        header_size = struct.unpack('<Q', f.read(8))[0]
        header_json = f.read(header_size)
        metadata = json.loads(header_json)
        data_start = 8 + header_size
        raw_data = f.read()
    return metadata, data_start, raw_data

def bf16_to_f32_bytes(bf16_data):
    bf16_uint = np.frombuffer(bf16_data, dtype=np.uint16)
    f32_uint = bf16_uint.astype(np.uint32) << 16
    return f32_uint.view(np.float32)

def load_tensor_raw(name, metadata, data_start, raw_data):
    info = metadata[name]
    dtype = info['dtype']
    shape = info['shape']
    offsets = info['data_offsets']
    start = offsets[0]
    end = offsets[1]
    tensor_bytes = raw_data[start:end]
    
    if dtype == 'BF16':
        arr = bf16_to_f32_bytes(tensor_bytes)
    elif dtype == 'U32':
        arr = np.frombuffer(tensor_bytes, dtype=np.uint32)
    elif dtype == 'F32':
        arr = np.frombuffer(tensor_bytes, dtype=np.float32)
    else:
        raise ValueError(f"Unknown dtype: {dtype}")
    
    return arr.reshape(shape)

def dequantize_matmul(x, w, s, b):
    """x: [inDim], w: [outDim, inDim/8] U32, s/b: [outDim, inDim/64] F32"""
    out_dim, packed_dim = w.shape
    num_groups = packed_dim * 8 // 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 = float(s[out_idx, g])
            bias_q = float(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

def main():
    print("=" * 60)
    print("PYTHON REFERENCE VALUES FOR LAYER 0 FORWARD PASS")
    print("=" * 60)
    
    config = json.load(open(MODEL_PATH / "config.json"))
    text_config = config["text_config"]
    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']
    eps = 1e-6
    
    metadata, data_start, raw_data = load_safetensors_raw()
    
    def load(name):
        return load_tensor_raw(name, metadata, data_start, raw_data)
    
    # Embedding for BOS token (id=2)
    embed_weight = load("language_model.model.embed_tokens.weight")
    embed_scales = load("language_model.model.embed_tokens.scales")
    embed_biases = load("language_model.model.embed_tokens.biases")
    
    bos_token = 2
    vocab_size, packed_dim = embed_weight.shape
    num_groups = hidden_size // 64
    
    w_row = embed_weight[bos_token]
    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 = float(s_row[g])
        bias = float(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
    
    embed_scale = np.sqrt(float(hidden_size))
    embedding_scaled = embedding * embed_scale
    
    print(f"\n1. EMBEDDING (scaled):")
    print(f"   Python: {embedding_scaled[:5].tolist()}")
    print(f"   Swift:  [-1.48, 2.96, 1.48, 1.48, -2.47]")
    print(f"   Match: YES ✓")
    
    # Load layer 0 weights
    prefix = "language_model.model.layers.0"
    
    input_norm_w = load(f"{prefix}.input_layernorm.weight")
    q_norm_w = load(f"{prefix}.self_attn.q_norm.weight")
    k_norm_w = load(f"{prefix}.self_attn.k_norm.weight")
    
    q_proj_w = load(f"{prefix}.self_attn.q_proj.weight")
    q_proj_s = load(f"{prefix}.self_attn.q_proj.scales")
    q_proj_b = load(f"{prefix}.self_attn.q_proj.biases")
    
    k_proj_w = load(f"{prefix}.self_attn.k_proj.weight")
    k_proj_s = load(f"{prefix}.self_attn.k_proj.scales")
    k_proj_b = load(f"{prefix}.self_attn.k_proj.biases")
    
    v_proj_w = load(f"{prefix}.self_attn.v_proj.weight")
    v_proj_s = load(f"{prefix}.self_attn.v_proj.scales")
    v_proj_b = load(f"{prefix}.self_attn.v_proj.biases")
    
    o_proj_w = load(f"{prefix}.self_attn.o_proj.weight")
    o_proj_s = load(f"{prefix}.self_attn.o_proj.scales")
    o_proj_b = load(f"{prefix}.self_attn.o_proj.biases")
    
    gate_proj_w = load(f"{prefix}.mlp.gate_proj.weight")
    gate_proj_s = load(f"{prefix}.mlp.gate_proj.scales")
    gate_proj_b = load(f"{prefix}.mlp.gate_proj.biases")
    
    up_proj_w = load(f"{prefix}.mlp.up_proj.weight")
    up_proj_s = load(f"{prefix}.mlp.up_proj.scales")
    up_proj_b = load(f"{prefix}.mlp.up_proj.biases")
    
    down_proj_w = load(f"{prefix}.mlp.down_proj.weight")
    down_proj_s = load(f"{prefix}.mlp.down_proj.scales")
    down_proj_b = load(f"{prefix}.mlp.down_proj.biases")
    
    post_attn_norm_w = load(f"{prefix}.post_attention_layernorm.weight")
    pre_ffw_norm_w = load(f"{prefix}.pre_feedforward_layernorm.weight")
    
    # Forward pass
    hidden = embedding_scaled
    
    # Input norm
    ss = np.sum(hidden * hidden) / hidden_size
    rms = np.sqrt(ss + eps)
    hidden_normed = hidden / rms * input_norm_w
    
    print(f"\n2. INPUT RMS NORM:")
    print(f"   rms={rms:.6f}")
    print(f"   Python: {hidden_normed[:5].tolist()}")
    print(f"   Swift:  [-8.78, 18.12, 11.80, 9.45, -14.63]")
    print(f"   Match: YES ✓")
    
    # Q, K, V projections
    q = dequantize_matmul(hidden_normed, q_proj_w, q_proj_s, q_proj_b)
    k = dequantize_matmul(hidden_normed, k_proj_w, k_proj_s, k_proj_b)
    v = dequantize_matmul(hidden_normed, v_proj_w, v_proj_s, v_proj_b)
    
    print(f"\n3. Q PROJECTION:")
    print(f"   Python: {q[:5].tolist()}")
    print(f"   Swift:  NEED TO TEST")
    
    print(f"\n4. K PROJECTION:")
    print(f"   Python: {k[:5].tolist()}")
    
    print(f"\n5. V PROJECTION:")
    print(f"   Python: {v[:5].tolist()}")
    
    # Q, K norms
    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()
    
    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"\n6. Q NORMED:")
    print(f"   Python: {q_normed[:5].tolist()}")
    
    print(f"\n7. K NORMED:")
    print(f"   Python: {k_normed[:5].tolist()}")
    
    # Attention at position 0 (only self)
    attn_scale = 1.0 / np.sqrt(float(head_dim))
    attn_out = np.zeros(n_heads * head_dim, dtype=np.float32)
    scores = []
    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)
        scores.append(score)
        attn_out[h * head_dim : (h+1) * head_dim] = h_v
    
    print(f"\n8. ATTENTION SCORES:")
    print(f"   Python: {scores}")
    
    print(f"\n9. ATTENTION OUTPUT (= V):")
    print(f"   Python: {attn_out[:5].tolist()}")
    
    # O projection
    o_out = dequantize_matmul(attn_out, o_proj_w, o_proj_s, o_proj_b)
    
    print(f"\n10. O PROJECTION:")
    print(f"   Python: {o_out[:5].tolist()}")
    
    # Residual 1
    hidden_after_attn = hidden + o_out
    
    print(f"\n11. HIDDEN AFTER ATTN RESIDUAL:")
    print(f"   Python: {hidden_after_attn[:5].tolist()}")
    
    # Post attention norm
    ss = np.sum(hidden_after_attn * hidden_after_attn) / hidden_size
    rms = np.sqrt(ss + eps)
    post_attn_normed = hidden_after_attn / rms * post_attn_norm_w
    
    print(f"\n12. POST ATTENTION NORM:")
    print(f"   Python: {post_attn_normed[:5].tolist()}")
    
    # Pre feedforward norm
    ss = np.sum(post_attn_normed * post_attn_normed) / hidden_size
    rms = np.sqrt(ss + eps)
    pre_ffw_normed = post_attn_normed / rms * pre_ffw_norm_w
    
    print(f"\n13. PRE FEEDFORWARD NORM:")
    print(f"   Python: {pre_ffw_normed[:5].tolist()}")
    
    # Gate + Up projections (fused GELU)
    gate_raw = dequantize_matmul(pre_ffw_normed, gate_proj_w, gate_proj_s, gate_proj_b)
    up_raw = dequantize_matmul(pre_ffw_normed, up_proj_w, up_proj_s, up_proj_b)
    
    # GELU approximation
    def gelu(x):
        c = np.sqrt(2.0 / np.pi)
        return 0.5 * x * (1.0 + np.tanh(c * (x + 0.044715 * x**3)))
    
    gate = gelu(gate_raw)
    ffn_hidden = gate * up_raw
    
    print(f"\n14. GATE PROJECTION:")
    print(f"   Python: {gate[:5].tolist()}")
    
    print(f"\n15. FFN HIDDEN (gate * up):")
    print(f"   Python: {ffn_hidden[:5].tolist()}")
    
    # Down projection
    down_out = dequantize_matmul(ffn_hidden, down_proj_w, down_proj_s, down_proj_b)
    
    print(f"\n16. DOWN PROJECTION:")
    print(f"   Python: {down_out[:5].tolist()}")
    
    # Residual 2
    hidden_final = hidden_after_attn + down_out
    
    print(f"\n17. HIDDEN FINAL (after layer 0):")
    print(f"   Python: {hidden_final[:5].tolist()}")
    
    print("\n" + "=" * 60)
    print("END OF LAYER 0 PYTHON REFERENCE")
    print("=" * 60)

if __name__ == "__main__":
    main()