kali20gakki/msAgent / skills/accuracy/uwc-verify
uwc-verify
Verified自动验证 veRL + Megatron-core + vLLM 训推权重转换一致性(UWC)的 5 项检查:Bridge P0/P2、传输 P2_sender/P5、data_ptr 覆盖、G0/G1 前向、dummy vs safetensors 差分。注入采集钩子、采集权重/指针/日志,生成 report.md / report.html / report.json。当用户要验证 RL 训练与推理权重同步、排查 rollout 输出乱码/漏同步、或复现《RL 权重一致性验证》文档时使用。
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 7.4 KB |
| references/reference.md | markdown | 10.9 KB |
| runtime/uwc/__init__.py | script | 373 B |
| runtime/uwc/collectors/__init__.py | script | 245 B |
| runtime/uwc/collectors/stage1_bridge_collector.py | script | 9.8 KB |
| runtime/uwc/collectors/stage2_transfer_collector.py | script | 2.9 KB |
| runtime/uwc/collectors/stage3_vllm_collector.py | script | 12.8 KB |
| runtime/uwc/common.py | script | 4.1 KB |
| runtime/uwc/comparator.py | script | 7.4 KB |
| runtime/uwc/fault.py | script | 8.4 KB |
| runtime/uwc/hooks.py | script | 26.2 KB |
| runtime/uwc/logits_comparator.py | script | 3.7 KB |
| runtime/uwc/name_normalizer.py | script | 7.3 KB |
| runtime/uwc/offline/__init__.py | script | 139 B |
| runtime/uwc/offline/forward_verify.py | script | 11.1 KB |
| runtime/uwc/offline/stage3_diff_test.py | script | 11.4 KB |
| runtime/uwc/offline/uwc_compare.py | script | 14.0 KB |
| runtime/uwc/offline/uwc_summary.py | script | 20.5 KB |
| scripts/uwc_verify.py | script | 44.4 KB |
| uwc.example.yaml | config | 3.4 KB |
runtime/uwc/comparator.py
View on GitHub"""Numeric comparison utilities for UWC weight dumps."""
from __future__ import annotations
import math
from typing import Iterable
NON_WEIGHT_SUFFIXES = (
"rotary_emb.inv_freq",
"rope_freqs",
"_extra_state",
)
def is_non_weight_key(name: str) -> bool:
return any(name.endswith(suffix) or suffix in name for suffix in NON_WEIGHT_SUFFIXES)
def _as_flat_float(tensor):
import torch
if not isinstance(tensor, torch.Tensor):
tensor = torch.as_tensor(tensor)
return tensor.detach().float().reshape(-1)
def tensor_stats(tensor) -> dict:
"""Return cheap statistics used in reports (works for empty tensors too)."""
import torch
if not isinstance(tensor, torch.Tensor):
tensor = torch.as_tensor(tensor)
flat = tensor.detach().float().reshape(-1)
if flat.numel() == 0:
return {"shape": list(tensor.shape), "dtype": str(tensor.dtype), "numel": 0}
finite = torch.isfinite(flat)
if finite.any():
values = flat[finite]
return {
"shape": list(tensor.shape),
"dtype": str(tensor.dtype),
"numel": int(flat.numel()),
"mean": float(values.mean().item()),
"std": float(values.std().item()) if values.numel() > 1 else 0.0,
"abs_max": float(values.abs().max().item()),
}
return {
"shape": list(tensor.shape),
"dtype": str(tensor.dtype),
"numel": int(flat.numel()),
"mean": "nan",
"std": "nan",
"abs_max": "nan",
}
def compare_tensors(a, b, atol: float = 1e-3, rtol: float = 0.0, exact: bool = False) -> dict:
"""Compare two tensors and return a JSON-serialisable result."""
import torch
if not isinstance(a, torch.Tensor):
a = torch.as_tensor(a)
if not isinstance(b, torch.Tensor):
b = torch.as_tensor(b)
result = {
"shape_a": list(a.shape),
"shape_b": list(b.shape),
"dtype_a": str(a.dtype),
"dtype_b": str(b.dtype),
"max_abs_diff": None,
"mean_abs_diff": None,
"mismatch": 0,
"numel": int(a.numel()),
"passed": False,
"reason": "",
}
if list(a.shape) != list(b.shape):
result["reason"] = f"shape mismatch: {list(a.shape)} vs {list(b.shape)}"
return result
af = _as_flat_float(a)
bf = _as_flat_float(b)
if af.numel() == 0:
result["passed"] = True
result["max_abs_diff"] = 0.0
result["mean_abs_diff"] = 0.0
return result
nan_a = torch.isnan(af)
nan_b = torch.isnan(bf)
nan_mismatch = int((nan_a != nan_b).sum().item())
finite = ~(nan_a | nan_b)
if finite.any():
diff = (af[finite] - bf[finite]).abs()
max_abs = float(diff.max().item())
mean_abs = float(diff.mean().item())
else:
diff = af.new_empty(0)
max_abs = 0.0 if nan_mismatch == 0 else float("nan")
mean_abs = 0.0 if nan_mismatch == 0 else float("nan")
if exact:
# Bitwise equality for the original dtype. NaN == NaN is treated as equal.
same = a.detach().cpu() == b.detach().cpu()
same = same | (torch.isnan(a.detach().cpu()) & torch.isnan(b.detach().cpu()))
mismatch = int((~same).sum().item())
threshold_desc = "exact"
else:
tol = atol + rtol * bf[finite].abs() if finite.any() else af.new_empty(0)
mismatch = int((diff > tol).sum().item()) if finite.any() else 0
mismatch += nan_mismatch
threshold_desc = f"atol={atol:g}" + (f", rtol={rtol:g}" if rtol else "")
result.update(
{
"max_abs_diff": max_abs,
"mean_abs_diff": mean_abs,
"mismatch": mismatch,
"nan_mismatch": nan_mismatch,
"threshold": threshold_desc,
"passed": mismatch == 0,
}
)
if mismatch:
result["reason"] = f"{mismatch}/{a.numel()} elements exceed {threshold_desc}"
return result
def _normalise_container(weights):
"""Accept dict[str, Tensor], list[(name, Tensor)] or a generator."""
if isinstance(weights, dict):
return dict(weights)
if hasattr(weights, "items"):
return dict(weights.items())
return {str(name): tensor for name, tensor in weights}
def compare_weight_dicts(
reference,
candidate,
atol: float = 1e-3,
rtol: float = 0.0,
exact: bool = False,
ignore_non_weight: bool = True,
name_normalizer=None,
worst_k: int = 20,
) -> dict:
"""Compare two name -> tensor mappings.
``name_normalizer`` is an optional callable applied to both key sets, e.g.
``uwc.name_normalizer.normalize_weights``. It must return either a dict or
a ``(dict, meta)`` tuple.
"""
ref = _normalise_container(reference)
cand = _normalise_container(candidate)
if name_normalizer is not None:
ref = name_normalizer(ref)
cand = name_normalizer(cand)
if isinstance(ref, tuple):
ref = ref[0]
if isinstance(cand, tuple):
cand = cand[0]
if ignore_non_weight:
ref = {k: v for k, v in ref.items() if not is_non_weight_key(k)}
cand = {k: v for k, v in cand.items() if not is_non_weight_key(k)}
missing = sorted(set(ref) - set(cand))
extra = sorted(set(cand) - set(ref))
common = sorted(set(ref) & set(cand))
details = []
mismatch_count = 0
num_weights = 0
global_max = 0.0
global_sum = 0.0
for key in common:
item = compare_tensors(ref[key], cand[key], atol=atol, rtol=rtol, exact=exact)
num_weights += 1
if item["mismatch"]:
mismatch_count += 1
if item["max_abs_diff"] is not None and not math.isnan(item["max_abs_diff"]):
global_max = max(global_max, item["max_abs_diff"])
global_sum += item["max_abs_diff"] * max(item["numel"], 1)
details.append({"name": key, **item})
details.sort(key=lambda item: (item["max_abs_diff"] is None, -(item["max_abs_diff"] or 0.0)))
worst = details[:worst_k]
passed = not missing and not extra and mismatch_count == 0
reasons = []
if missing:
reasons.append(f"{len(missing)} keys missing in candidate")
if extra:
reasons.append(f"{len(extra)} unexpected keys in candidate")
if mismatch_count:
reasons.append(f"{mismatch_count}/{num_weights} weights mismatch")
return {
"status": "pass" if passed else "fail",
"reference_keys": len(ref),
"candidate_keys": len(cand),
"compared_keys": num_weights,
"mismatched_keys": mismatch_count,
"missing_keys": missing[:50],
"extra_keys": extra[:50],
"missing_count": len(missing),
"extra_count": len(extra),
"max_abs_diff": global_max,
"mean_abs_diff": (global_sum / sum(max(item["numel"], 1) for item in details)) if details else 0.0,
"threshold": f"exact" if exact else f"atol={atol:g}" + (f", rtol={rtol:g}" if rtol else ""),
"worst": worst,
"reasons": reasons,
}
def filter_weights(weights, keep: Iterable[str] | None = None, drop: Iterable[str] | None = None):
keep_set = set(keep) if keep is not None else None
drop_set = set(drop) if drop is not None else None
result = {}
for name, tensor in _normalise_container(weights).items():
if keep_set is not None and name not in keep_set:
continue
if drop_set is not None and name in drop_set:
continue
result[name] = tensor
return result
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |