Correct Hybrid FLOP Calculation

[Phlip79]

Summary

This PR splits the HybridModel TFLOPs/MFU accounting fix out of the GPT-OSS HybridModel migration.

HybridModel commonly represents one logical decoder block as multiple physical hybrid layers, for example *E for attention followed by MoE. Bridge's generic FLOP accounting needs to count those physical symbols as the corresponding logical attention, MLP, and MoE work. Without that, GPT-OSS-style *E layouts can report roughly doubled TFLOPs/MFU even when runtime throughput is unchanged.

Changes:

• route configs with hybrid_layer_pattern through hybrid FLOP accounting
• use MCore's canonical get_hybrid_layer_counts(), parse_hybrid_pattern(), and Symbols helpers for HybridModel pattern handling
• split hybrid attention FLOPs into token-linear projection work and core-attention work so seqlen_squared_sum is honored
• account for sliding-window attention on physical hybrid attention layers
• add parity tests comparing equivalent transformer and HybridModel physical patterns

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[Phlip79]

/ok to test e6f0afe

[claude]

Review: Correct Hybrid FLOP Calculation

The math for splitting attention FLOPs into projection (linear in seq_len) and core-attention (quadratic) terms looks correct and is consistent with the transformer path. The SWA accounting properly mirrors the existing transformer SWA logic. Good parity test coverage.

Two items to clarify:

1. Import path divergence: flop_utils.py now imports from megatron.core.models.hybrid.hybrid_layer_allocation, while the rest of the codebase (hybrid_provider.py, mlm_compat/model.py) still uses megatron.core.ssm.mamba_hybrid_layer_allocation. If this is the new canonical MCore location, the old sites should be updated in a follow-up.

2. GPT-OSS test assertion: The new assertion provider.is_hybrid_model is True in test_gpt_oss_bridges.py -- the GPT-OSS bridge does not set this field, and it is not in CONFIG_MAPPING. If this depends on a concurrent PR or MCore change, this test will fail in isolation.

No perf tests impacted.

[Phlip79]

/ok to test 9551726