Open
Conversation
…on on Neuron BioReason-Pro combines ESM3-small (~1.4B) protein encoder with Qwen3-4B backbone via NxDI for protein function prediction. Includes 8 runtime patches enabling inputs_embeds passthrough in NxDI 0.8.0, plus full integration test suite validated on trn2.3xlarge (SDK 2.28).
When sweeping multiple batch sizes, compiled artifacts (model.pt, weights/, neuron_config.json) were saved to the model weights directory, causing subsequent batch sizes to load the first compiled model instead of recompiling. The new compiled_model_path parameter directs compiled artifacts to a separate directory per batch size while still reading HF weights from model_path.
…tok/s at DP=4 BS=4 - Add predict_batch() method for batched inputs_embeds generation - Fix ESM3 tokenizer mask_token incompatibility with transformers >= 4.47 (mask_token returns None in ESM3 >= 3.0.4, breaking tokenize_sequence) - Update dp_launcher worker to collect batches when batch_size > 1 - DP=4 BS=4 benchmark: 299.9 tok/s aggregate (2.1x GPU A10G peak) 40 proteins, 64,568 tokens, 215.3s wall time, 0 errors - Update README with DP+batching results and compatibility matrix
- Add measured DP sweep: LNC=2 DP=4 BS=4/8/16 and LNC=1 DP=8 BS=1/4/8 - Peak: 537.3 tok/s (LNC=2 DP=4 BS=16) vs H100 best 522.6 tok/s = 1.03x - LNC=2 DP=4 confirmed optimal (1.52x faster than LNC=1 DP=8) - Update compatibility matrix with all validated configs - All numbers measured, no projections
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
inputs_embedspassthrough through NxDI compiled modelDataParallelRunnerfor multi-core inferenceBenchmark Results (trn2.3xlarge, SDK 2.28)
Single-Core Performance
Data Parallelism + Batching
GPU Comparison (A10G)
Files
src/modeling_bioreason.py— BioReasonPipeline with ESM3 encoder, embedding injection, predict/predict_batchsrc/patch_nxdi_embeds.py— 8 V3b patches for inputs_embeds passthroughsrc/dp_launcher.py— DataParallelRunner for multi-core inference with batchingtest/integration/test_model.py— 5 integration teststest/integration/benchmark_dp.py— Data-parallel benchmark scriptKey Technical Details