Parity Testing for Megatron Bridge

This skill provides the decision framework for choosing the right verification tool and interpreting results. For the full model onboarding workflow (which includes parity testing as milestones 1 and 2), see the add-model-support skill.

Quick Decision: Which Tool to Run

All tools live under examples/conversion/.

3-Level Test Strategy

Level 1: State Dict Round-Trip (exact match)

The fastest and most fundamental check. If mappings can't perfectly round-trip weights, nothing else will work.

# Single-GPU round-trip
uv run python examples/conversion/hf_megatron_roundtrip.py \
    --hf-model-id <org>/<model>

# Multi-GPU with TP=2
uv run python -m torch.distributed.run --nproc_per_node=2 \
    examples/conversion/hf_megatron_roundtrip_multi_gpu.py \
    --hf-model-id <org>/<model> --tp 2

# Multi-GPU with PP=2
uv run python -m torch.distributed.run --nproc_per_node=2 \
    examples/conversion/hf_megatron_roundtrip_multi_gpu.py \
    --hf-model-id <org>/<model> --pp 2

Expected: Every weight shows "Matches Original: checkmark". Any "X" means the param mapping has an error.

Tolerance: Exact match (max_diff == 0.0). Round-trip conversions are pure tensor reshaping — no floating-point arithmetic is involved.

For programmatic verification inside scripts, use the built-in verifier:

from megatron.bridge.models.conversion.utils import weights_verification_table
weights_verification_table(bridge, hf_pretrained, megatron_model)

Level 2: Forward-Pass Correlation (GPU / bfloat16)

After round-trip passes, verify that the converted model produces strongly correlated logits and the same next-token prediction. This is a functional correlation gate, not a claim of bitwise-identical floating-point arithmetic.

# Compare logits (loads both HF and Megatron models)
uv run python -m torch.distributed.run --nproc_per_node=2 \
    examples/conversion/compare_hf_and_megatron/compare.py \
    --hf_model_path <org>/<model> --tp 2 \
    --prompt "The capital of France is"

Expected: Matching next-token predictions and cosine similarity at least 0.99. Equivalently, cosine distance (1 - cosine_similarity) must be at most 1%. Record numeric maximum and mean absolute logit differences for diagnosis, but do not use either absolute difference as a pass/fail guard. A fixed absolute threshold is not scale- or ULP-aware across BF16 logits and implementations.

For large models that OOM compare.py (which loads both models), use text generation as a separate inference sanity check:

uv run python -m torch.distributed.run --nproc_per_node=2 \
    examples/conversion/hf_to_megatron_generate_text.py \
    --hf_model_path <org>/<model> --tp 2 \
    --prompt "The capital of France is" --max_new_tokens 50

Generation does not measure logit correlation and therefore cannot by itself verify the manual forward-pass item.

Level 3: Training Parity (optional)

Verify that a few training steps produce decreasing loss. This catches gradient computation issues that forward-pass tests miss. Use a toy model with 2 layers and small dimensions. See the functional test pattern in the add-model-support skill (Milestone 3, Phase 6).

Tolerance Table

Keep stricter cosine or absolute-difference thresholds as optional diagnostic targets when investigating arithmetic differences. Do not use them to downgrade a model-support card after the correlation gate passes.

Transformer Engine normalization

Using Transformer Engine LayerNorm/RMSNorm can change floating-point operation ordering relative to the HF or PyTorch reference. Treat that as a forward-correlation question, not a weight-conversion failure: state-dict round-trip must remain exact, while forward parity uses the cosine and next-token gates above. Do not replace an available TE affine norm solely to reduce a report-only absolute logit difference. Conversely, do not add affine parameters to weightless reference norms just to use TE; that changes the architecture and invalidates exact mapping coverage.

Comparison Utilities

These functions are useful when writing custom verification scripts or debugging failures. They are not part of the Bridge library — copy them into your script as needed.

import torch


def compare_tensors(a, b, name=""):
    """Compare two tensors and report similarity metrics."""
    max_diff = (a - b).abs().max().item()
    mean_diff = (a - b).abs().mean().item()
    cos_sim = torch.nn.functional.cosine_similarity(
        a.flatten().float(), b.flatten().float(), dim=0,
    ).item()
    print(f"{name}: max_diff={max_diff:.6e}, mean_diff={mean_diff:.6e}, cosine_sim={cos_sim:.8f}")
    return max_diff, mean_diff, cos_sim


def compare_state_dicts(sd_a, sd_b, prefix=""):
    """Compare two state dicts key-by-key, reporting per-parameter differences."""
    keys_a, keys_b = set(sd_a.keys()), set(sd_b.keys())
    missing, extra = keys_a - keys_b, keys_b - keys_a
    if missing:
        print(f"{prefix}Missing keys: {sorted(missing)}")
    if extra:
        print(f"{prefix}Extra keys: {sorted(extra)}")
    max_diffs = {}
    for key in sorted(keys_a & keys_b):
        diff = (sd_a[key].float() - sd_b[key].float()).abs().max().item()
        if diff > 0:
            max_diffs[key] = diff
            print(f"{prefix}{key}: max_diff={diff:.6e}")
    if not max_diffs and not missing and not extra:
        print(f"{prefix}All {len(keys_a & keys_b)} parameters match exactly.")
    return missing, extra, max_diffs

Debugging Workflow

When a parity test fails, follow this sequence:

1. Run single-GPU round-trip — if this fails, the mapping itself is

wrong. Check the mapping_registry() in the bridge file.

1. If single-GPU passes but multi-GPU fails — the TP/PP scatter/gather

is wrong. Compare the TP=1 result against each TP shard. See the nccl-contiguous-tensors skill for NCCL-specific issues.

1. If round-trip passes but forward correlation fails — the weights loaded

correctly, but the runtime architectures may differ. A failure means the next token differs or cosine similarity is below 0.99; a large absolute difference alone is diagnostic, not a failure. Check provider_bridge() config mapping (normalization, activation, RoPE, etc.).

1. Use the debugging script template from the add-model-support skill

to inspect runtime vs safetensors key naming and bridge config mapping.

For the full catalog of pitfalls (QKV interleaving, MoE fused exports, tied embeddings, FP8 dequantization, TE LayerNorm aliases, etc.), see the Pitfalls section of the add-model-support skill.

Code Anchors