MarkBase Admin
ac75faa0cc
CI / build-and-test (push) Has been cancelled
- E4B-MarkBase model (42 layers, 4.4GB) loaded successfully - All Phase 1-6 tests passed (model loading, forward pass, vision/audio towers, token generation, performance) - All stress tests passed (5/5 in 127.6s) - Concurrent inference - Memory stress (67.5 tok/s, 0 NaN) - Continuous generation - Batch processing - Long-running stability - Swift Metal inference engine with multimodal support
245 lines
8.9 KiB
Python
245 lines
8.9 KiB
Python
#!/usr/bin/env python3
|
|
"""Debug E4B inference using MLX to compare with Swift Metal."""
|
|
|
|
import mlx.core as mx
|
|
import numpy as np
|
|
import json
|
|
from pathlib import Path
|
|
import struct
|
|
|
|
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_bytes(bf16_data):
|
|
"""Convert raw BF16 bytes to float32 numpy array."""
|
|
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_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 load_tensor_raw(name, metadata, data_start, raw_data):
|
|
"""Load tensor from raw safetensors 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}")
|
|
|
|
# Reshape to proper dimensions
|
|
return arr.reshape(shape)
|
|
|
|
def debug_inference():
|
|
print("=" * 50)
|
|
print("MLX/NumPy E4B Debug Inference")
|
|
print("=" * 50)
|
|
|
|
config = load_config()
|
|
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']
|
|
|
|
print(f"\nConfig: hidden={hidden_size}, head_dim={head_dim}")
|
|
print(f" n_heads={n_heads}, n_kv_heads={n_kv_heads}")
|
|
|
|
# Load safetensors
|
|
metadata, data_start, raw_data = load_safetensors_raw()
|
|
|
|
def load(name):
|
|
return load_tensor_raw(name, metadata, data_start, raw_data)
|
|
|
|
print("\n--- Loading weights ---")
|
|
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")
|
|
|
|
print(f"embed_weight: shape={embed_weight.shape}, dtype={embed_weight.dtype}")
|
|
print(f"embed_scales: shape={embed_scales.shape}, dtype={embed_scales.dtype}")
|
|
|
|
# Dequantize BOS embedding
|
|
bos_token = 2
|
|
vocab_size, packed_dim = embed_weight.shape
|
|
num_groups = hidden_size // 64
|
|
|
|
print(f"\n--- Dequantizing BOS (token={bos_token}) ---")
|
|
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
|
|
|
|
print(f"Embedding: min={embedding.min():.6f}, max={embedding.max():.6f}")
|
|
print(f" first 20: {embedding[:20].tolist()}")
|
|
|
|
# Scale
|
|
embed_scale = np.sqrt(float(hidden_size))
|
|
embedding_scaled = embedding * embed_scale
|
|
print(f"\nScaled (x{embed_scale:.2f}): min={embedding_scaled.min():.6f}, max={embedding_scaled.max():.6f}")
|
|
print(f" first 20: {embedding_scaled[:20].tolist()}")
|
|
|
|
# Layer 0
|
|
print("\n--- Layer 0 ---")
|
|
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")
|
|
|
|
print(f"input_norm: {input_norm_w.shape}, q_norm: {q_norm_w.shape}")
|
|
print(f"q_proj: w={q_proj_w.shape}, s={q_proj_s.shape}")
|
|
|
|
# Forward
|
|
def dequantize_matmul(x, w, s, b):
|
|
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
|
|
|
|
hidden = embedding_scaled
|
|
eps = 1e-6
|
|
|
|
# Input norm
|
|
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" first 5: {hidden_normed[:5].tolist()}")
|
|
|
|
# Q, K, V
|
|
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"\nQ: min={q.min():.6f}, max={q.max():.6f}")
|
|
print(f" first 5: {q[:5].tolist()}")
|
|
print(f"\nK: min={k.min():.6f}, max={k.max():.6f}")
|
|
print(f" first 5: {k[:5].tolist()}")
|
|
print(f"\nV: min={v.min():.6f}, max={v.max():.6f}")
|
|
print(f" first 5: {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"\nQ normed: min={q_normed.min():.6f}, max={q_normed.max():.6f}")
|
|
print(f" first 5: {q_normed[:5].tolist()}")
|
|
print(f"\nK normed: min={k_normed.min():.6f}, max={k_normed.max():.6f}")
|
|
print(f" first 5: {k_normed[:5].tolist()}")
|
|
|
|
# Attention (pos 0)
|
|
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"\nAttention scores: {scores}")
|
|
print(f"Attn out: min={attn_out.min():.6f}, max={attn_out.max():.6f}")
|
|
print(f" first 5: {attn_out[:5].tolist()}")
|
|
|
|
# O proj
|
|
o_out = dequantize_matmul(attn_out, o_proj_w, o_proj_s, o_proj_b)
|
|
print(f"\nO out: min={o_out.min():.6f}, max={o_out.max():.6f}")
|
|
print(f" first 5: {o_out[:5].tolist()}")
|
|
|
|
# Residual
|
|
hidden_after = hidden + o_out
|
|
print(f"\nAfter attn: min={hidden_after.min():.6f}, max={hidden_after.max():.6f}")
|
|
print(f" first 5: {hidden_after[:5].tolist()}")
|
|
|
|
print("\n" + "=" * 50)
|
|
print("KEY VALUES TO COMPARE:")
|
|
print("=" * 50)
|
|
print(f"\nEmbedding scaled: {embedding_scaled[:5].tolist()}")
|
|
print(f"After input norm: {hidden_normed[:5].tolist()}")
|
|
print(f"Q projection: {q[:5].tolist()}")
|
|
print(f"Q normed: {q_normed[:5].tolist()}")
|
|
print(f"Attention out: {attn_out[:5].tolist()}")
|
|
print(f"O projection: {o_out[:5].tolist()}")
|
|
print(f"Hidden after attn: {hidden_after[:5].tolist()}")
|
|
|
|
if __name__ == "__main__":
|
|
debug_inference() |