Results from neuronxcc-2.28.7631.0+8f7b0cf5

README.md

@@ -5,6 +5,7 @@ tags:
 - neuron
 - trainium
 - transformers
+- sentence-transformers
 configs:
 - config_name: default
   data_files:
@@ -17,8 +18,9 @@ configs:
 Which Hugging Face Hub models run unmodified on AWS Trainium, through the
 native PyTorch backend
 [`torch-neuronx`](https://github.com/aws-neuron/torch-neuronx), for each
-`neuronx-cc` and library version tested. The only library tested so far is
-[`transformers`](https://github.com/huggingface/transformers).
+`neuronx-cc` and library version tested. Tested libraries are
+[`transformers`](https://github.com/huggingface/transformers) and
+[`sentence-transformers`](https://github.com/huggingface/sentence-transformers).
 
 To look up a model, use the
 [Space](https://huggingface.co/spaces/aws-neuron/torch-neuronx-model-support).
@@ -101,6 +103,65 @@ run 100 models, skipping those whose test script has not changed since their
 last run. Quantized checkpoints and models that need `trust_remote_code` are
 left out. A model that leaves the list keeps the results it already has.
 
+### sentence-transformers
+
+Each run scores one query against four documents using `SentenceTransformer`,
+`CrossEncoder`, or `MultiVectorEncoder`. SentenceTransformer encodes the query
+and documents and computes their similarity; CrossEncoder predicts a score for
+each query/document pair; MultiVectorEncoder compares token embeddings with
+MaxSim. Saved retrieval prompts and any model-specific task settings are recorded
+in the run script. All three paths use batch size 4 and the same input:
+
+```python
+query = "Which planet is known as the Red Planet?"
+documents = [
+    "Venus is often called Earth's twin because of its similar size and proximity.",
+    "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
+    "Jupiter, the largest planet in our solar system, has a prominent red spot.",
+    "Saturn, famous for its rings, is sometimes mistaken for the Red Planet.",
+]
+```
+
+The model runs on CPU for reference, then on one logical Neuron device using
+native eager PyTorch. The optional compile sweep uses `model.compile()` with the
+Neuron backend for model forward calls; tokenization and similarity scoring
+remain outside that wrapper. It applies the same checks and records the mode and
+number of compiled graphs. Compile results are kept locally because the published
+schema does not distinguish execution modes.
+
+FP32 is the default, with BF16 for large checkpoints or
+following an allocation failure. SDPA is preferred, with eager attention retried
+when unsupported by the checkpoint. Each script records the checkpoint revision,
+Sentence Transformers source commit, compiler version, and final settings.
+Checkpoints that require custom code load with `trust_remote_code=True`.
+
+`PASS` requires four finite Neuron scores, Mars ranked first, CPU agreement at
+the same precision (`atol=rtol=0.001`), a stable repeat with the same tolerance,
+and a completed device-memory capture. `FAIL(output)` means a ranking, parity,
+or repeat check failed after execution. A ranking failure can also occur on CPU;
+it does not by itself indicate a Neuron defect. Loading and forward failures
+record their stage and traceback. Scores retain document order and are not
+comparable between models. This checks one retrieval example, not general
+retrieval quality or throughput.
+
+Memory is the highest observed sum of runtime HBM categories, sampled every
+100 ms from before model transfer until the final inference is synchronized.
+Capacity comes from the device properties. Values use GiB (bytes / 2**30),
+despite the Space's historical `Memory (GB)` label. Short peaks between samples
+may be missed; incomplete captures leave memory empty. Wall time includes
+imports, loading, the CPU reference, Neuron compilation and inference, the repeat,
+and monitoring. It is not per-query latency. `runner` names the encoder class;
+`neuron_cores=1` denotes one logical device, with no tensor parallelism.
+
+Selection takes the exact top 100 models from each Hub listing tagged
+`sentence-transformers`, by Trending and Downloads, and tests their deduplicated
+union. Sparse encoders, GGUF/MLX and quantized checkpoints, and unrelated
+classification tasks are excluded; excluded or failing entries are not replaced.
+The saved selection retains both ranks and checkpoint revisions. `selection_rank`
+is the Trending rank; Downloads ranks appear in notes and raw results. Scripts,
+logs, scores, memory samples, and prior attempts are saved under
+`validation/sentence-transformers/` in the dataset.
+
 ## Layout
 
 There is one file per Neuron target (`trn2`, `trn3`) and tested

data/trn2/neuronxcc-2.28.7631.0+8f7b0cf5/transformers-5.22.1.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:aa40ac1984637767bd9f772e4c45c7726aced8ac3fada3daa83d1cf7cf6d07b5
-size 1375415
+oid sha256:1165863729e70731b547f7f3f5cdb81a6a38eeae54a605c63b63a43b62f8c26e
+size 1377073