Huvu/gemma4 e4b tp sp#5508
Draft
huvunvidia wants to merge 15 commits into
Draft
Conversation
…+ I-b) Gemma4TransformerConfig subclasses TransformerConfig with per-layer get_config_for_layer (256/512 kv_channels, 1e4/1e6 rotary_base, sliding window) plus forward-only flags (softcap, embed scaling, PLE dims). The three primitive modules are bitwise-faithful ports of HF Gemma4: - Gemma4RMSNorm: fp32 pow(-0.5), plain weight, with_scale flag. - Gemma4RotaryEmbedding: sliding full-rotary + full proportional RoPE. - Gemma4PLE: per_layer_inputs with bf16-rounded embed scales. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…pecs (I-c) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per-layer config dropped gemma4-only fields (hidden_size_per_layer_input, layer_types, etc.) and passed non-base rotary_base into TransformerConfig. Carry all non-base fields + derived layer_types onto the per-layer config; rope theta is handled by Gemma4RotaryEmbedding by layer_type, not config.
build_module returns a FunctionType unchanged rather than invoking it, so build_module(gemma4_rms_norm_scaleless_builder, ...) left self.v_layernorm as the builder function itself, which crashed at forward time (apply_module: 'function is not a subclass of torch.nn.Module'). Call the builder directly, matching how the base attention/layer build q/k/input layernorms from their LayerNormBuilder slots. This makes the I-c suite fully green (V6_local, kv_bus, V6_te all PASS). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Thin entrypoint mirroring pretrain_gpt.py that builds Gemma4Model with Gemma4TransformerConfig and the gemma4 local/TE layer specs. Reuses the GPT data pipeline, batch logic, loss, and forward step. The final-logit softcap and sqrt(H) embedding scaling live inside Gemma4Model.forward and reach the training loss path. Includes a self-contained nvrx __version__ shim applied before megatron import. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Base TransformerConfig defaults normalization='LayerNorm', so the TE spec's TENorm was building a mean-subtracting te.pytorch.LayerNorm instead of RMSNorm, causing large TE-spec divergence (V4 logits max-abs ~36, per-layer ~90). The local spec was unaffected (hard-wires Gemma4RMSNorm). Forcing RMSNorm in __post_init__ drops TE per-layer divergence to bf16-noise (~0-3).
…ing) The bias_dropout_add path runs in-place in eval/no-grad mode, which aliases the post-norm output tensor across the attention/MLP sub-blocks and corrupts the sandwich-norm result. Gemma4 sublayers have no bias and dropout=0, so add the residual explicitly: h = residual + post_norm(sublayer(pre_norm(h))). Mirrors HF Gemma4TextDecoderLayer exactly. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
per_layer_input_gate used gather_output=False, which at TP>1 would shard the gate to ple_dim/TP while the threaded per_layer_input (p_i) stays full-width, so gate * p_i mismatches (e.g. 32 vs 256 at TP=8). Gather the gate so the PLE sub-block runs replicated. No-op at TP=1 (numerically identical); does not affect weight sharding or checkpoint load. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: ataghibakhsh <ataghibakhsh@nvidia.com>
Make Gemma 4 E4B correct under tensor + sequence parallelism (TP=2/4/8 +SP), training-only, parity with the verified TP=1 oracle. - PLE SP layout (gemma4_model.py, gemma4_layer.py): unify the per-layer-input gate/projection as a standard Column(gather_output=False)->Row(input_is_parallel= True) pair and shard per_layer_inputs on the PLE dim via scatter_to_tensor_model_parallel_region. Gemma4Model.forward gathers decoder_input (SP-guarded, tensor_parallel_output_grad=False to avoid a TP gradient overcount) to recompute full-sequence PLE while the residual stream stays sequence-sharded. Replaces the prior gather_output=True hack; correct at TP=1 / TP>1 noSP / TP>1+SP. - Norm grad-tag (gemma4_norm.py): mark weighted Gemma4RMSNorm weights sequence_parallel so their replicated-but-partial grads all-reduce across TP under SP (mirrors TENorm). - Config hardening (pretrain_gemma4.py): force qk_layernorm / add_bias_linear / add_qkv_bias and assert GQA num_query_groups for correct TP>num_query_groups sharding. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: ataghibakhsh <ataghibakhsh@nvidia.com>
V0-V7 verification tooling (lives outside megatron/, supporting the TP+SP work): - tp_fwd.py: TP/SP-parameterized forward (vocab-gather, SP-aware per-layer capture, --pad-seq-to for SP sequence-divisibility), mlm_env.py worktree path shim. - compare_logits.py: parity verdict (greedy + logits/per-layer bands, --real-len). - diag_pos.py: per-position top-2 logit diagnostic (near-tie vs bug). - tp_reshard.py: V7 dist-ckpt reshard round-trip. - v0_dryrun / tp_parity / v_sp2 / v_sp4 / v_sp8 / v_sp_all / v6_* / v7_reshard sbatch. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: ataghibakhsh <ataghibakhsh@nvidia.com>
…TP/SP harness) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: ataghibakhsh <ataghibakhsh@nvidia.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: ataghibakhsh <ataghibakhsh@nvidia.com>
…n+parity) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: ataghibakhsh <ataghibakhsh@nvidia.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Signed-off-by: ataghibakhsh <ataghibakhsh@nvidia.com>
Contributor
|
This PR has been automatically converted to draft because all PRs must start as drafts. When you are ready for review, click Ready for Review to begin the review process. This will:
See the contribution guide for more details. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do ?
Issue tracking
For PRs from open-source community contributors:
Linked issue:
Contribution process
Pre-checks
Code review
Feel free to message or comment @NVIDIA/mcore-oncall to help accelerate your merge into main. The less complex your PR is, the faster it will be approved and merged!
All PRs start as draft. If you open a non-draft PR, it will be automatically converted to draft.
Step 1: Mark PR as "Ready for Review"
.github/CODEOWNERS.Final Review might get declined if these requirements are not fulfilled.
Step 2: Final Review
For PRs that change
megatron/core, once all expert reviewers have approved, theFinal Reviewlabel is applied automatically and final reviewers are assigned.For PRs outside
megatron/core, this step is skipped.Step 3: Approved
Once all required reviewers have approved, the
Approvedlabel is applied automatically.Merge
Any member of mcore-engineers will be able to merge your PR.