File size: 7,891 Bytes
54152e6 | 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 | """Load the complete sharded Mage-Flow NVFP4 transformer component."""
from __future__ import annotations
from contextlib import ExitStack
import json
from pathlib import Path
from typing import Any
import torch
import torch.nn as nn
from safetensors import safe_open
from packed_artifact import (
assign_tensor_by_name,
build_target_specs,
instantiate_mage_transformer_on_meta,
materialize_mage_rope_tensor_attributes,
set_child_module,
unregistered_meta_tensor_attribute_names,
)
from torch_ops_native import (
PackedNvfp4LinearNativeOp,
initialize_native_sm120_op,
)
class StandardCheckpointError(RuntimeError):
pass
def fail(message: str) -> None:
raise StandardCheckpointError(message)
def read_object(path: Path) -> dict[str, Any]:
try:
value = json.loads(path.read_text(encoding="utf-8"))
except (OSError, json.JSONDecodeError) as exc:
fail(f"cannot read JSON object {path}: {exc}")
if not isinstance(value, dict):
fail(f"expected a JSON object: {path}")
return value
def _component_path(component_dir: Path, relative: str) -> Path:
path = (component_dir / relative).resolve()
if not path.is_relative_to(component_dir):
fail(f"checkpoint index path escapes transformer component: {relative}")
if not path.is_file():
fail(f"checkpoint shard is missing: {relative}")
return path
def _quantized_keys(module_key: str) -> dict[str, str]:
return {
"packed_weight": f"{module_key}.packed_weight",
"weight_scales": f"{module_key}.weight_scales",
"weight_scale": f"{module_key}.weight_scale",
"bias": f"{module_key}.bias",
}
def load_standard_native_transformer(
repo_root: str | Path,
device: torch.device,
) -> tuple[nn.Module, dict[str, Any]]:
"""Load a complete standard-layout component without a BF16 base download."""
repo_root = Path(repo_root).resolve()
component_dir = (repo_root / "transformer").resolve()
if not component_dir.is_relative_to(repo_root) or not component_dir.is_dir():
fail("repository has no transformer component")
if device.type != "cuda":
fail("the native resident transformer requires a CUDA destination")
if not initialize_native_sm120_op(allow_python_schema_fallback=False):
fail("compiled native SM120 torch op did not load")
config = read_object(component_dir / "config.json")
quant_config = config.get("quantization_config")
if not isinstance(quant_config, dict):
fail("transformer config has no quantization_config")
if quant_config.get("quant_method") != "mage_flow_nvfp4":
fail(
"unexpected transformer quantization method: "
f"{quant_config.get('quant_method')!r}"
)
if quant_config.get("quant_algo") != "NVFP4":
fail("transformer config does not declare NVFP4")
depth = int(config.get("depth", 0))
target_specs = build_target_specs(depth)
expected_targets = [spec.module_key for spec in target_specs]
if quant_config.get("targets") != expected_targets:
fail("transformer quantization target list is not canonical")
metadata = read_object(component_dir / "nvfp4_metadata.json")
if metadata.get("artifact_kind") != (
"mage_flow_transformer_mlp_nvfp4_resident_v1"
):
fail("unexpected transformer NVFP4 metadata kind")
non_target_keys = metadata.get("non_target_keys")
if not isinstance(non_target_keys, list) or not all(
isinstance(key, str) for key in non_target_keys
):
fail("transformer NVFP4 metadata has no non-target key list")
index = read_object(
component_dir / "diffusion_pytorch_model.safetensors.index.json"
)
weight_map = index.get("weight_map")
if not isinstance(weight_map, dict) or not all(
isinstance(key, str) and isinstance(value, str)
for key, value in weight_map.items()
):
fail("transformer checkpoint has no valid weight map")
quantized_keys = {
key
for spec in target_specs
for key in _quantized_keys(spec.module_key).values()
}
expected_keys = set(non_target_keys) | quantized_keys
actual_keys = set(weight_map)
if actual_keys != expected_keys:
missing = sorted(expected_keys - actual_keys)
unexpected = sorted(actual_keys - expected_keys)
fail(
"transformer checkpoint key coverage mismatch; "
f"missing={missing[:1]}, unexpected={unexpected[:1]}"
)
original_target_weights = {spec.weight_key for spec in target_specs}
leaked = sorted(actual_keys & original_target_weights)
if leaked:
fail(f"BF16 target weight leaked into quantized checkpoint: {leaked[0]}")
shard_names = sorted(set(weight_map.values()))
with ExitStack() as stack:
handles = {
name: stack.enter_context(
safe_open(
_component_path(component_dir, name),
framework="pt",
device="cpu",
)
)
for name in shard_names
}
def tensor(key: str) -> torch.Tensor:
try:
handle = handles[weight_map[key]]
except KeyError:
fail(f"tensor is absent from checkpoint index: {key}")
if key not in handle.keys():
fail(f"tensor is absent from its declared shard: {key}")
return handle.get_tensor(key)
model = instantiate_mage_transformer_on_meta(repo_root)
for spec in target_specs:
original = model.get_submodule(spec.module_key)
if not isinstance(original, nn.Linear):
fail(
f"expected target {spec.module_key} to be nn.Linear, "
f"found {type(original).__name__}"
)
keys = _quantized_keys(spec.module_key)
replacement = PackedNvfp4LinearNativeOp(
in_features=int(original.in_features),
out_features=int(original.out_features),
packed_weight=tensor(keys["packed_weight"]).to(device),
weight_scales=tensor(keys["weight_scales"]).to(device),
weight_scale=tensor(keys["weight_scale"]).to(device),
bias=tensor(keys["bias"]).to(device),
)
set_child_module(model, spec.module_key, replacement)
loaded_non_targets: list[str] = []
for key in non_target_keys:
assign_tensor_by_name(model, key, tensor(key).to(device))
loaded_non_targets.append(key)
materialized = materialize_mage_rope_tensor_attributes(model)
meta_parameters = [
name for name, value in model.named_parameters() if value.is_meta
]
meta_buffers = [
name for name, value in model.named_buffers() if value.is_meta
]
unregistered_meta = unregistered_meta_tensor_attribute_names(model)
if meta_parameters or meta_buffers or unregistered_meta:
fail(
"standard transformer loader left unresolved meta tensors: "
f"{(meta_parameters + meta_buffers + unregistered_meta)[:4]}"
)
report = {
"layout": "huggingface_sharded_component",
"checkpoint_shard_count": len(shard_names),
"checkpoint_tensor_count": len(actual_keys),
"loaded_non_target_tensor_count": len(loaded_non_targets),
"loaded_quantized_projection_count": len(target_specs),
"bf16_target_weight_reads": 0,
"meta_parameter_names": meta_parameters,
"meta_buffer_names": meta_buffers,
"materialized_unregistered_tensor_attribute_names": materialized,
}
return model.eval().requires_grad_(False), report
__all__ = [
"StandardCheckpointError",
"load_standard_native_transformer",
]
|