File size: 8,496 Bytes
48ecd01 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 | #!/usr/bin/env python3
"""Migrate checkpoint from separate Q/K/V projections to fused QKV.
Usage:
python3 scripts/migrate_qkv_checkpoint.py <checkpoint_dir>
Migrates both model.pt AND optimizer.pt:
- model.pt: q_proj/k_proj/v_proj weights β qkv_proj weight
- optimizer.pt: exp_avg/exp_avg_sq states fused, param indices re-mapped
The concatenation order is [Q ; K ; V] along the output (dim-0) axis,
which matches the split in MultiHeadAttention.forward:
q, k, v = qkv.split([_q_dim, _kv_dim, _kv_dim], dim=-1)
Optimizer layout (group 0 = weight_decay, per layer Γ 28):
[i*6+0] q_proj.weight [3072, 3072]
[i*6+1] k_proj.weight [1024, 3072]
[i*6+2] v_proj.weight [1024, 3072]
[i*6+3] out_proj.weight [3072, 3072]
[i*6+4] fc1_weight [16384, 3072]
[i*6+5] fc2_weight [3072, 8192]
After fusion: indices 0,1,2 β single qkv_proj β 4 params per layer.
"""
import sys
import torch
from pathlib import Path
N_LAYERS = 28
OLD_PARAMS_PER_LAYER = 6 # q, k, v, out, fc1, fc2
NEW_PARAMS_PER_LAYER = 4 # qkv, out, fc1, fc2
def migrate_model(state: dict) -> dict:
"""Fuse Q/K/V projection weights into QKV in model state dict."""
new_state: dict = {}
layers_done: set = set()
for key, val in state.items():
if ".q_proj." not in key and ".k_proj." not in key and ".v_proj." not in key:
new_state[key] = val
continue
if ".q_proj." not in key:
continue
prefix = key.rsplit(".", 2)[0]
suffix = key.rsplit(".", 1)[-1]
tag = (prefix, suffix)
if tag in layers_done:
continue
layers_done.add(tag)
q_key = f"{prefix}.q_proj.{suffix}"
k_key = f"{prefix}.k_proj.{suffix}"
v_key = f"{prefix}.v_proj.{suffix}"
missing = [k for k in (q_key, k_key, v_key) if k not in state]
if missing:
raise KeyError(f"Expected keys not found in checkpoint: {missing}")
q_w, k_w, v_w = state[q_key], state[k_key], state[v_key]
fused = torch.cat([q_w, k_w, v_w], dim=0)
fused_key = f"{prefix}.qkv_proj.{suffix}"
new_state[fused_key] = fused
print(f" Fused {fused_key}: {list(fused.shape)}"
f" (q={list(q_w.shape)}, k={list(k_w.shape)}, v={list(v_w.shape)})")
leaked = [k for k in new_state if ".q_proj." in k or ".k_proj." in k or ".v_proj." in k]
if leaked:
raise RuntimeError(f"BUG: old projection keys still present: {leaked}")
return new_state
def migrate_optimizer(opt_state: dict) -> dict:
"""Fuse optimizer states for Q/K/V β QKV and re-index parameters.
The optimizer has 2 param groups:
Group 0 (weight_decay): 168 = 28 layers Γ 6 (q,k,v,out,fc1,fc2)
Group 1 (no weight_decay): 58 = norms + embedding
We fuse q,k,v entries in group 0 (indices i*6+0,1,2 β one entry per layer).
Group 0 shrinks from 168 to 112 (28 layers Γ 4 params).
Group 1 stays at 58. Total: 170.
"""
old_state = opt_state["state"]
old_groups = opt_state["param_groups"]
group0_count = len(old_groups[0]["params"])
expected_g0 = N_LAYERS * OLD_PARAMS_PER_LAYER
if group0_count != expected_g0:
raise ValueError(
f"Group 0 has {group0_count} params, expected {expected_g0}. "
f"Cannot auto-detect QKV layout."
)
# Validate shapes for first layer
shapes = []
for j in range(OLD_PARAMS_PER_LAYER):
idx = old_groups[0]["params"][j]
shapes.append(list(old_state[idx]["exp_avg"].shape))
expected_shapes = [[3072, 3072], [1024, 3072], [1024, 3072],
[3072, 3072], [16384, 3072], [3072, 8192]]
if shapes != expected_shapes:
raise ValueError(
f"Layer 0 shapes {shapes} don't match expected {expected_shapes}. "
f"Cannot auto-detect QKV layout."
)
print(f" Shape validation passed for layer 0.")
new_state_entries = {}
new_idx = 0
# --- Group 0: fuse q/k/v per layer ---
for layer_i in range(N_LAYERS):
base = layer_i * OLD_PARAMS_PER_LAYER
q_opt_idx = old_groups[0]["params"][base + 0]
k_opt_idx = old_groups[0]["params"][base + 1]
v_opt_idx = old_groups[0]["params"][base + 2]
q_entry = old_state[q_opt_idx]
k_entry = old_state[k_opt_idx]
v_entry = old_state[v_opt_idx]
# Fuse QKV
fused_entry = {"step": q_entry["step"]}
for field in ["exp_avg", "exp_avg_sq"]:
if field in q_entry:
fused_entry[field] = torch.cat(
[q_entry[field], k_entry[field], v_entry[field]], dim=0
)
new_state_entries[new_idx] = fused_entry
if layer_i == 0:
print(f" Layer 0 QKV fused: exp_avg {list(fused_entry['exp_avg'].shape)}")
new_idx += 1
# Copy remaining params (out, fc1, fc2)
for offset in [3, 4, 5]:
opt_idx = old_groups[0]["params"][base + offset]
new_state_entries[new_idx] = old_state[opt_idx]
new_idx += 1
new_group0_count = new_idx # should be N_LAYERS * NEW_PARAMS_PER_LAYER = 112
print(f" Group 0: {group0_count} β {new_group0_count} params")
# --- Group 1: copy as-is (norms, embedding β no QKV) ---
group1_count = len(old_groups[1]["params"])
for j in range(group1_count):
opt_idx = old_groups[1]["params"][j]
if opt_idx in old_state:
new_state_entries[new_idx] = old_state[opt_idx]
new_idx += 1
print(f" Group 1: {group1_count} β {group1_count} params (unchanged)")
# Build new param_groups
new_groups = []
g0 = {k: v for k, v in old_groups[0].items() if k != "params"}
g0["params"] = list(range(0, new_group0_count))
new_groups.append(g0)
g1 = {k: v for k, v in old_groups[1].items() if k != "params"}
g1["params"] = list(range(new_group0_count, new_group0_count + group1_count))
new_groups.append(g1)
total = new_group0_count + group1_count
print(f" Total: {len(old_state)} β {total} optimizer params")
return {"state": new_state_entries, "param_groups": new_groups}
def migrate(ckpt_dir: Path) -> None:
model_path = ckpt_dir / "model.pt"
opt_path = ckpt_dir / "optimizer.pt"
if not model_path.exists():
raise FileNotFoundError(f"model.pt not found in {ckpt_dir}")
# --- Model migration ---
print(f"[1/2] Migrating model weights from {model_path} ...")
state = torch.load(model_path, map_location="cpu", weights_only=True)
has_old = any(".q_proj." in k for k in state)
has_new = any(".qkv_proj." in k for k in state)
if has_new and not has_old:
print(" Model already migrated. Skipping.")
elif has_old:
new_model_state = migrate_model(state)
torch.save(new_model_state, model_path)
print(f" Model saved.")
else:
raise RuntimeError("Model state has neither q_proj nor qkv_proj keys!")
# --- Optimizer migration ---
if opt_path.exists():
print(f"\n[2/2] Migrating optimizer states from {opt_path} ...")
opt = torch.load(opt_path, map_location="cpu", weights_only=True)
# Check if already migrated
total_params = sum(len(pg["params"]) for pg in opt["param_groups"])
expected_old = N_LAYERS * OLD_PARAMS_PER_LAYER + 58 # 168 + 58 = 226
expected_new = N_LAYERS * NEW_PARAMS_PER_LAYER + 58 # 112 + 58 = 170
if total_params == expected_old:
opt_backup = ckpt_dir / "optimizer.pt.backup_pre_qkv"
if not opt_backup.exists():
torch.save(opt, opt_backup)
print(f" Backup: {opt_backup}")
new_opt = migrate_optimizer(opt)
torch.save(new_opt, opt_path)
print(f" Optimizer saved.")
elif total_params == expected_new:
print(f" Optimizer already migrated ({total_params} params). Skipping.")
else:
print(f" [WARN] Unexpected param count {total_params} "
f"(expected old={expected_old} or new={expected_new}). "
f"Deleting optimizer.pt β optimizer will restart fresh.")
opt_path.unlink()
else:
print("\n[2/2] No optimizer.pt found. Optimizer will restart fresh.")
print("\nMigration complete!")
if __name__ == "__main__":
if len(sys.argv) != 2:
print(__doc__)
sys.exit(1)
migrate(Path(sys.argv[1]))
|