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)

Check tensor coverage and mapping behavior before interpreting forward results. Pure layout transformations must round-trip weights exactly. Investigate arithmetic parameter transformations separately from missing or misplaced weights.

# 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 mapped by pure layout transformations shows "Matches Original: checkmark". Investigate every "X" before assigning a cause.

Tolerance: Exact match (max_diff == 0.0) for pure reshaping, permutation, and sharding. If a mapping performs arithmetic or changes dtype, report that transformation and its measured effect explicitly. Use the controlled module-swap procedure below before classifying differences as numerical; do not silently relax the exact-match check.

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: unchanged parameter representations must still round-trip exactly; investigate arithmetic parameter transformations as described above. 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:

Before classifying any discrepancy as numerical, follow the controlled module-swap procedure in skills/adding-model-support/SKILL.md. Magnitude, BF16 precision, or a suspected normalization kernel is not sufficient evidence. Fix confirmed bugs; preserve reusable production blocks for confirmed numerical differences and retain the diagnostic evidence.

1. Run single-GPU round-trip — inspect the mapping_registry(), layouts, dtypes, and arithmetic parameter transformations. Missing tensors or wrong shapes indicate conversion defects; investigate value differences before assigning a cause.
2. If single-GPU passes but multi-GPU fails — inspect TP/PP/EP distribution, gathering, and checkpoint metadata. Compare the TP=1 result against each TP shard. See the nccl-contiguous-tensors skill for NCCL-specific issues.
3. If round-trip passes but forward correlation fails — the tested roundtrip recovered the source tensors. Verify runtime parameter placement and effective values; inverse mapping errors can cancel in a roundtrip. Record cosine similarity, absolute differences, token decisions, and continuations; use task-specific acceptance criteria. A high cosine alone does not explain a changed token or establish numerical equivalence. Check the bridge's config mapping in hf_config_to_model_config_kwargs(), or in provider_bridge() for families that have not migrated (normalization, activation, RoPE, etc.).
4. 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