[None][feat] Eagle: PostNorm and multilayer options #9233
Conversation
IzzyPutterman
changed the title
|
/bot run |
📝 WalkthroughWalkthroughThese changes add conditional architectural paths to the Eagle3 speculative model based on configuration flags. A new Changes
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes
Pre-merge checks and finishing touches❌ Failed checks (2 warnings, 1 inconclusive)
✨ Finishing touches
🧪 Generate unit tests (beta)
Tip 📝 Customizable high-level summaries are now available in beta!You can now customize how CodeRabbit generates the high-level summary in your pull requests — including its content, structure, tone, and formatting.
Example instruction:
Note: This feature is currently in beta for Pro-tier users, and pricing will be announced later. Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Nitpick comments (5)
tensorrt_llm/_torch/models/modeling_speculative.py (5)
33-37: Confirm QKV projection behavior fornext_layer_regular=True vs False
Eagle3Attentionnow only overridesself.qkv_projwhenself._next_layer_regularis False, otherwise it uses the baseAttentionQKV (with hidden-size input) instead of the EAGLE3-style 2×hidden-size input andFUSED_QKV_LINEARweight mode. This changes both the expected input shape and the weight loading/quantization behavior between “regular” and non-regular layers.Please double‑check that:
- All call sites feeding this attention respect the two possible input shapes (hidden vs 2×hidden).
- The weight mapping correctly handles the different
qkv_projshapes/modes between regular and non‑regular layers, especially when sharing weights with the target model.If this dichotomy is intentional, consider a brief comment here explaining the two regimes (
_next_layer_regularvs non‑regular) for future readers.Also applies to: 57-70
79-90: Verify_next_layer_regularlogic and embedding concatenation semantics
_next_layer_regularis computed as:self._next_layer_regular = (eagle_config.get("next_layer_regular", True) and not is_first_layer) or eagle_config.get( "eh_proj_before_attn", False)and then:
input_layernormand the[embeds, hidden_states]concatenation are only used whennot self._next_layer_regular.Eagle3Attention’s QKV override is likewise gated onnot self._next_layer_regular.This implies:
- The first decoder layer always uses the EAGLE3 2×hidden-size attention path when
eh_proj_before_attnis False, regardless ofnext_layer_regular.- When
eh_proj_before_attnis True, all layers behave as “regular” (no concat, regular QKV).That seems plausible but is quite subtle. Please verify that this matches the intended architecture (especially whether configs are ever expected to make the first layer “regular”), and that for all eagle_config combinations:
embedsandhidden_stateshave matching last-dimension sizes when concatenated.- The chosen attention QKV shape matches the concatenated vs non‑concatenated input.
If this logic is correct, a short comment near the
_next_layer_regularassignment clarifying these cases would help avoid future regressions.Also applies to: 106-109, 130-132
80-80: Fix__init__return type annotation
Eagle3DecoderLayer.__init__is annotated as returningTuple[torch.Tensor, torch.Tensor], but__init__must always returnNonein Python. This will confuse type checkers and IDEs.Consider changing the signature to:
- ) -> Tuple[torch.Tensor, torch.Tensor]: + ) -> None:
163-171: Pre‑attentioneh_proj_before_attnpath: check shapes and module choiceThe new
eagle_config["eh_proj_before_attn"]path introduces:
self.enorm: RMSNorm(hidden_size)self.eh_proj: nn.Linear(2 * hidden_size, hidden_size, dtype=config.torch_dtype)- Forward logic that normalizes
inputs_embeds, concatenates withhidden_states, then projects back tohidden_size.Points to verify:
- For all supported configs, both
inputs_embedsandhidden_statesindeed have last dimensionconfig.hidden_sizeat this point, so the concat produces exactly2 * hidden_sizeas expected.- This pre‑attention fusion is performed exactly once where intended (i.e., not re‑applied inside the decoder layers due to
_next_layer_regulargating).Also, most of this file uses the custom
Linearwrapper for tensor parallelism and quantization, whileeh_projusesnn.Lineardirectly. If this projection is performance‑critical or should participate in TP/quant, consider switching it toLinearwith appropriateWeightsLoadingConfig; otherwise, documenting that this is intentionally a non‑TP/non‑quantized helper would clarify the design.Also applies to: 204-212, 258-262
178-179: Clarify_return_hidden_post_normcontract and call‑site expectationsThe new
_return_hidden_post_normflag changes theEagle3DraftModel.forwardreturn from:hidden_states, hidden_states_to_save = self.norm(hidden_states, residual) if self._return_hidden_post_norm: return hidden_states, hidden_states return hidden_states, hidden_states_to_saveCallers now may receive either
(post_norm, post_norm)or(post_norm, something_else)depending on config.Please ensure that:
- All current call sites consuming the second tensor (if any) can handle both behaviors and are keyed off the same
eagle_config["return_hidden_post_norm"]flag.- The semantics of
hidden_states_to_savewhen_return_hidden_post_normis False remain consistent with whatSpecMetadata.get_hidden_states()/ the pyexecutor expect.A brief docstring or comment describing what “post_norm” vs “to_save” represents, and when configs should enable this flag, would make the behavior much clearer.
Also applies to: 280-284
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
tensorrt_llm/_torch/models/modeling_speculative.py(10 hunks)
🧰 Additional context used
🧠 Learnings (1)
📓 Common learnings
Learnt from: venkywonka
Repo: NVIDIA/TensorRT-LLM PR: 6029
File: .github/pull_request_template.md:45-53
Timestamp: 2025-08-27T17:50:13.264Z
Learning: For PR templates in TensorRT-LLM, avoid suggesting changes that would increase developer overhead, such as converting plain bullets to mandatory checkboxes. The team prefers guidance-style bullets that don't require explicit interaction to reduce friction in the PR creation process.
🧬 Code graph analysis (1)
tensorrt_llm/_torch/models/modeling_speculative.py (1)
tensorrt_llm/_torch/modules/linear.py (4)
Linear(1880-2105)TensorParallelMode(50-62)WeightsLoadingConfig(45-47)WeightMode(35-41)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Pre-commit Check
|
PR_Github #24801 [ run ] triggered by Bot. Commit: |
|
PR_Github #24801 [ run ] completed with state |
|
/bot run |
|
PR_Github #24925 [ run ] triggered by Bot. Commit: |
|
PR_Github #24925 [ run ] completed with state |
|
/bot run |
|
PR_Github #24997 [ run ] triggered by Bot. Commit: |
|
/bot run |
|
PR_Github #25142 [ run ] triggered by Bot. Commit: |
|
PR_Github #25142 [ run ] completed with state |
| if self.num_layers > 1: | ||
| self.midlayer = nn.ModuleList([ | ||
| Eagle3DecoderLayer(model_config, start_layer_idx + i) | ||
| Eagle3DecoderLayer(model_config, start_layer_idx + i, i == 0) |
There was a problem hiding this comment.
nitpick, I prefer this for readability
Eagle3DecoderLayer(model_config, start_layer_idx + i, is_first_layer=i == 0)| # we expect that to happen outside the model definition. This helps us | ||
| # avoid data-dependent control flow and gives us better CUDA graph | ||
| # coverage. | ||
| # ideally,we expect that to happen outside the model definition. This |
There was a problem hiding this comment.
| # ideally,we expect that to happen outside the model definition. This | |
| # ideally, we expect that to happen outside the model definition. This |
IzzyPutterman
force-pushed
the
iputterman/eagle-options-mtp-main
branch
from
8719fd9 to
8ca8f52
Compare
|
/bot run |
|
PR_Github #25218 [ run ] triggered by Bot. Commit: |
|
PR_Github #25218 [ run ] completed with state |
|
/bot run |
|
PR_Github #25220 [ run ] triggered by Bot. Commit: |
|
PR_Github #25220 [ run ] completed with state |
Signed-off-by: Izzy Putterman <[email protected]>
IzzyPutterman
force-pushed
the
iputterman/eagle-options-mtp-main
branch
from
8ca8f52 to
f18df49
Compare
|
/bot run |
|
PR_Github #25227 [ run ] triggered by Bot. Commit: |
|
PR_Github #25227 [ run ] completed with state |
|
/bot run |
|
PR_Github #25248 [ run ] triggered by Bot. Commit: |
|
PR_Github #25248 [ run ] completed with state |
|
/bot run |
|
PR_Github #25315 [ run ] triggered by Bot. Commit: |
3 participants
Summary by CodeRabbit
New Features
Refactor
Description
Test Coverage
PR Checklist
Please review the following before submitting your PR:
PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
Update tava architecture diagram if there is a significant design change in PR.
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.
GitHub Bot Help
/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...Provide a user friendly way for developers to interact with a Jenkins server.
Run
/bot [-h|--help]to print this help message.See details below for each supported subcommand.
run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]Launch build/test pipelines. All previously running jobs will be killed.
--reuse-test (optional)pipeline-id(OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.--disable-reuse-test(OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.--disable-fail-fast(OPTIONAL) : Disable fail fast on build/tests/infra failures.--skip-test(OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.--stage-list "A10-PyTorch-1, xxx"(OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.--gpu-type "A30, H100_PCIe"(OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.--test-backend "pytorch, cpp"(OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.--only-multi-gpu-test(OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.--disable-multi-gpu-test(OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.--add-multi-gpu-test(OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.--post-merge(OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx"(OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".--detailed-log(OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.--debug(OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in thestage-listparameter to access the appropriate container environment. Note: Does NOT update GitHub check status.For guidance on mapping tests to stage names, see
docs/source/reference/ci-overview.mdand the
scripts/test_to_stage_mapping.pyhelper.kill
killKill all running builds associated with pull request.
skip
skip --comment COMMENTSkip testing for latest commit on pull request.
--comment "Reason for skipping build/test"is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.reuse-pipeline
reuse-pipelineReuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.