File size: 13,723 Bytes
e50ecc3 58d30dd e50ecc3 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 |
#!/usr/bin/env python3
# βββββββββββββββββββ make local repo override any wheel ββββββββββββββββββββ
import sys, os; sys.path.insert(0, os.path.abspath("."))
# βββββββββββββββββββ Flash-Attention + CUDA stubs ββββββββββββββββββββββββββ
import types, torch, torch.nn.functional as F, importlib.machinery as im
flash_pkg = types.ModuleType("flash_attn"); flash_pkg.__spec__ = im.ModuleSpec("flash_attn", loader=None, is_package=True); flash_pkg.__path__=[]
sys.modules["flash_attn"] = flash_pkg
fa = types.ModuleType("flash_attn.flash_attn_interface"); fa.__spec__ = im.ModuleSpec("flash_attn.flash_attn_interface", loader=None)
def _sdpa(qkv,*_,causal=False,**__): q,k,v = qkv.unbind(1); q,k,v = (t.unsqueeze(0) for t in(q,k,v)); return F.scaled_dot_product_attention(q,k,v,is_causal=causal).squeeze(0)
for s in ("flash_attn_unpadded_qkvpacked_func","flash_attn_unpadded_kvpacked_func","flash_attn_varlen_qkvpacked_func","flash_attn_varlen_kvpacked_func"): setattr(fa, s, _sdpa)
sys.modules["flash_attn.flash_attn_interface"] = fa; flash_pkg.flash_attn_interface = fa
pad = types.ModuleType("flash_attn.bert_padding"); pad.__spec__ = im.ModuleSpec("flash_attn.bert_padding", loader=None)
pad.pad_input = lambda x,*a,**k:(x,None); pad.unpad_input = lambda x,*a,**k:x
sys.modules["flash_attn.bert_padding"] = pad; flash_pkg.bert_padding = pad
if not torch.cuda.is_available():
torch.cuda.is_available=lambda:False
torch.cuda.get_device_capability=lambda dev=None:(0,0)
torch.cuda.current_device=lambda:0
torch.cuda.get_device_properties=lambda dev=None:types.SimpleNamespace(major=0,minor=0)
import importlib.metadata as _im
if "flash_attn" not in _im.packages_distributions():
rv, rd = _im.version, _im.distribution
_im.version = lambda p:"0.0.0" if p=="flash_attn" else rv(p)
_im.distribution = lambda p:types.SimpleNamespace(version="0.0.0") if p=="flash_attn" else rd(p)
# βββββββββββββββββββ std imports βββββββββββββββββββββββββββββββββββββββββββ
from pathlib import Path
import argparse, json, shutil
from huggingface_hub import hf_hub_download
from transformers import AutoConfig
from gr00t.model.gr00t_n1 import GR00T_N1_5
# βββββββββββββββββββ helpers βββββββββββββββββββββββββββββββββββββββββββββββ
def patched_cfg():
p = hf_hub_download("nvidia/GR00T-N1.5-3B", "config.json")
d = json.load(open(p))
if d.get("model_type") != "gr00t_n1_5":
d["model_type"] = "gr00t_n1_5"
patched = Path(p).with_name("config_patched.json")
patched.write_text(json.dumps(d)); return str(patched)
return p
def build_blank():
cfg = AutoConfig.from_pretrained(patched_cfg(),
trust_remote_code=True,
local_files_only=True)
cfg.backbone_cfg.update(dict(tune_llm=True)) # enable L-tower
cfg.backbone_cfg.pop("checkpoint_path", None)
cfg.backbone_cfg.pop("use_pretrained", None)
cfg.action_head_cfg.pop("checkpoint_path", None)
torch.manual_seed(0)
return GR00T_N1_5(cfg, local_model_path="") # random weights
def maybe_add_lm_head(model):
"""Ensure lm_head is properly initialized with weights"""
# Navigate to the language model
lm = model.backbone.eagle_model.language_model
# Get dimensions from embed_tokens
embed_tokens = lm.model.embed_tokens
vocab_size = embed_tokens.num_embeddings
hidden_size = embed_tokens.embedding_dim
print(f"Embedding dimensions: vocab_size={vocab_size}, hidden_size={hidden_size}")
# Expected shape based on architecture: [151680, 2048]
if vocab_size != 151680 or hidden_size != 2048:
print(f"β οΈ Warning: Unexpected dimensions. Expected vocab=151680, hidden=2048")
# Check if lm_head exists
if hasattr(lm, "lm_head"):
print(f"lm_head attribute exists: {lm.lm_head is not None}")
# Even if lm_head exists, it might not have weights properly initialized
# Just replace it with a properly initialized one
print("Creating new lm_head with proper initialization...")
else:
print("lm_head attribute missing, creating...")
# Create a new lm_head with proper initialization
# Note: nn.Linear uses (in_features, out_features), so it's (hidden_size, vocab_size)
new_lm_head = torch.nn.Linear(hidden_size, vocab_size, bias=False)
# Initialize weights with normal distribution (std=0.02 is standard for LM heads)
torch.nn.init.normal_(new_lm_head.weight, mean=0.0, std=0.02)
# Convert to bfloat16 to match backbone
new_lm_head.weight.data = new_lm_head.weight.data.to(torch.bfloat16)
# Replace the lm_head
lm.lm_head = new_lm_head
print(f"β Created lm_head: Linear({hidden_size}, {vocab_size}, bias=False)")
print(f" Weight shape: {lm.lm_head.weight.shape}")
print(f" Weight dtype: {lm.lm_head.weight.dtype}")
print(f" Parameters: {lm.lm_head.weight.numel() / 1e6:.1f}M")
def set_mixed(model):
"""Set mixed precision: backbone in bf16, action head in fp32"""
for n,p in model.named_parameters():
if n.startswith("backbone.") or "lm_head" in n:
p.data = p.data.to(torch.bfloat16)
else:
p.data = p.data.to(torch.float32)
def copy_tokenizer(out):
for f in ("tokenizer.json","tokenizer_config.json","vocab.txt","special_tokens_map.json"):
try: shutil.copy(hf_hub_download("nvidia/GR00T-N1.5-3B", f), out/f)
except Exception: pass
def diagnose_model(model):
"""Print diagnostic info about the model"""
print("\nModel diagnostics:")
total_params = sum(p.numel() for p in model.parameters())
print(f" Total params: {total_params/1e6:,.0f}M")
# Check for key components
has_lm_head = False
lm_head_params = 0
lm_head_location = None
for name, param in model.named_parameters():
if "lm_head" in name:
has_lm_head = True
lm_head_params += param.numel()
lm_head_location = name
print(f" Has lm_head: {'β' if has_lm_head else 'β'}")
if has_lm_head:
print(f" lm_head params: {lm_head_params/1e6:,.0f}M")
print(f" lm_head location: {lm_head_location}")
# Check if the params are actually counted in the total
lm = model.backbone.eagle_model.language_model
if hasattr(lm, 'lm_head') and lm.lm_head is not None:
actual_params = lm.lm_head.weight.numel()
print(f" lm_head actual params: {actual_params/1e6:,.0f}M")
print(f" lm_head weight shape: {lm.lm_head.weight.shape}")
print(f" lm_head weight dtype: {lm.lm_head.weight.dtype}")
def validate_model_architecture(model):
"""Validate model against the architecture specification"""
print("\n" + "="*60)
print("ARCHITECTURE VALIDATION")
print("="*60)
# Expected architecture based on the spec
expected_shapes = {
# Key layers to check - using actual parameter names with .weight suffix
"backbone.eagle_model.language_model.lm_head.weight": (151680, 2048),
"backbone.eagle_model.language_model.model.embed_tokens.weight": (151680, 2048),
"backbone.eagle_model.language_model.model.norm.weight": (2048,),
"backbone.eagle_model.mlp1.0.weight": (2048, 1152),
"backbone.eagle_model.mlp1.0.bias": (2048,),
"action_head.position_embedding.weight": (1024, 1536), # Fixed: added .weight
"action_head.vlln.weight": (2048,),
"action_head.vlln.bias": (2048,),
}
errors = []
warnings = []
# Get all parameters
param_dict = dict(model.named_parameters())
# Debug: print actual action_head parameter names to see the pattern
action_head_params = [name for name in param_dict.keys() if name.startswith("action_head.position")]
if action_head_params:
print("\nFound position embedding parameters:")
for name in action_head_params[:5]:
print(f" {name}: {param_dict[name].shape}")
# Check key shapes
for name, expected_shape in expected_shapes.items():
if name in param_dict:
actual_shape = tuple(param_dict[name].shape)
if actual_shape != expected_shape:
errors.append(f"Shape mismatch for {name}: expected {expected_shape}, got {actual_shape}")
else:
print(f"β {name}: {actual_shape}")
else:
errors.append(f"Missing parameter: {name}")
# Check dtypes
dtype_issues = []
for name, param in param_dict.items():
if name.startswith("backbone."):
if param.dtype != torch.bfloat16:
dtype_issues.append(f"{name}: expected bfloat16, got {param.dtype}")
elif name.startswith("action_head."):
if param.dtype != torch.float32:
dtype_issues.append(f"{name}: expected float32, got {param.dtype}")
if dtype_issues:
warnings.extend(dtype_issues[:5]) # Only show first 5
# Count parameters by component
component_params = {
"backbone": 0,
"action_head": 0,
"other": 0
}
for name, param in param_dict.items():
count = param.numel()
if name.startswith("backbone."):
component_params["backbone"] += count
elif name.startswith("action_head."):
component_params["action_head"] += count
else:
component_params["other"] += count
# Special check for lm_head
lm_head_found = False
lm_head_params = 0
for name, param in param_dict.items():
if "lm_head" in name:
lm_head_found = True
lm_head_params += param.numel()
# Report results
print("\nValidation Results:")
print(f" Errors: {len(errors)}")
print(f" Warnings: {len(warnings)}")
if errors:
print("\nβ ERRORS:")
for error in errors:
print(f" - {error}")
if warnings:
print("\nβ οΈ WARNINGS (showing first 5):")
for warning in warnings[:5]:
print(f" - {warning}")
if len(warnings) > 5:
print(f" ... and {len(warnings) - 5} more")
print("\nπ Parameter Summary:")
total = sum(component_params.values())
print(f" Total: {total/1e6:,.1f}M")
print(f" Backbone: {component_params['backbone']/1e6:,.1f}M")
print(f" Action Head: {component_params['action_head']/1e6:,.1f}M")
if component_params['other'] > 0:
print(f" Other: {component_params['other']/1e6:,.1f}M")
print(f"\n lm_head found: {'β' if lm_head_found else 'β'}")
if lm_head_found:
print(f" lm_head params: {lm_head_params/1e6:.1f}M (expected: 311.1M)")
# Expected totals based on NVIDIA model
expected_total = 2724 # Million params
actual_total = total / 1e6
diff = actual_total - expected_total
print(f"\n Expected total: {expected_total}M")
print(f" Actual total: {actual_total:.1f}M")
print(f" Difference: {diff:+.1f}M")
if abs(diff) < 1: # Within 1M params
print("\nβ
Model architecture matches expected specification!")
else:
print("\nβ Model architecture does NOT match specification!")
return len(errors) == 0
# βββββββββββββββββββ main ββββββββββββββββββββββββββββββββββββββββββββββββββ
def main(device: str, out_dir: str):
print("="*60)
print("Creating blank GR00T-N1.5-3B model")
print("="*60)
model = build_blank()
# Add diagnostics before adding lm_head
print("\nBefore adding lm_head:")
diagnose_model(model)
maybe_add_lm_head(model)
# Add diagnostics after adding lm_head
print("\nAfter adding lm_head:")
diagnose_model(model)
set_mixed(model)
model = model.to(device)
# Validate against architecture spec
validate_model_architecture(model)
out = Path(out_dir).expanduser(); out.mkdir(parents=True, exist_ok=True)
print(f"\nSaving model to {out}...")
model.save_pretrained(out, max_shard_size="2GB")
copy_tokenizer(out)
(out/"README.md").write_text("Random GR00T-N1.5-3B | backbone bf16 | action_head fp32 | Apache-2.0\n")
# Final summary
print("\n" + "="*60)
print("FINAL SUMMARY")
print("="*60)
print(f"β
Saved blank model ({sum(p.numel() for p in model.parameters())/1e6:,.0f}M params) β {out}")
print(f"β
Model has lm_head with {model.backbone.eagle_model.language_model.lm_head.weight.numel()/1e6:.1f}M params")
print(f"β
Ready for training with Apache-2.0 license")
# βββββββββββββββββββ CLI βββββββββββββββββββββββββββββββββββββββββββββββββββ
if __name__ == "__main__":
ap = argparse.ArgumentParser()
ap.add_argument("--device", default="cpu")
ap.add_argument("--out_dir", default="DolphinGR00T-N1.5-3B-Zero")
args = ap.parse_args(); main(args.device, args.out_dir) |