Results from neuronxcc-2.28.7631.0+8f7b0cf5

README.md

@@ -5,7 +5,6 @@ tags:
 - neuron
 - trainium
 - transformers
-- sentence-transformers
 configs:
 - config_name: default
   data_files:
@@ -18,9 +17,8 @@ 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. Tested libraries are
-[`transformers`](https://github.com/huggingface/transformers) and
-[`sentence-transformers`](https://github.com/huggingface/sentence-transformers).
+`neuronx-cc` and library version tested. The only library tested so far is
+[`transformers`](https://github.com/huggingface/transformers).
 
 To look up a model, use the
 [Space](https://huggingface.co/spaces/aws-neuron/torch-neuronx-model-support).
@@ -103,104 +101,6 @@ 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
-
-The 2026-11-03 sweep covers the top 100 models in each Hub listing tagged
-`sentence-transformers`: [Trending](https://huggingface.co/models?library=sentence-transformers&sort=trending)
-and [Downloads](https://huggingface.co/models?library=sentence-transformers&sort=downloads).
-There are 164 distinct checkpoints, with 36 appearing in both lists. The exact
-ordered lists, checkpoint revisions, and capture time are frozen in
-`validation/sentence-transformers/2026-11-03-top100/`. Every eligible checkpoint
-was tested afresh; failures were not replaced with passing models. SparseEncoder,
-non-native quantized formats, and unrelated classification tasks have explicit
-`EXCLUDED` records. `selection_rank` records the snapshot's Trending rank;
-Downloads ranks are in the raw results and row notes. The Space's Trending and
-Downloads views use the live Hub ordering, which can change after this snapshot.
-
-| Snapshot | PASS | Output check failed | Load/forward failed | Excluded | Total |
-|---|---:|---:|---:|---:|---:|
-| Trending | 71 | 10 | 9 | 10 | 100 |
-| Downloads | 80 | 6 | 12 | 2 | 100 |
-| Unique models | 119 | 15 | 18 | 12 | 164 |
-
-All 33 failed checkpoints were rerun after access was granted to the two gated
-repositories. Their outcome categories are unchanged, but the current errors and
-measurements are from these fresh runs. `litert-community/embeddinggemma-300m`
-contains TFLite exports rather than a native Transformers checkpoint and fails
-model loading; `google/embeddinggemma-300m` already passes. `nvidia/NV-Embed-v1`
-now reaches its required custom code (`trust_remote_code=True`), which imports
-`MISTRAL_INPUTS_DOCSTRING`, unavailable in the tested Transformers 5.22.1 stack.
-Neither current failure is an access error. The comparison is recorded in
-`validation/sentence-transformers/2026-11-03-top100/failed-rerun-1.json`;
-previous scripts, logs, results, and captured memory samples are preserved under
-`prior_runs/failed-rerun-1/` in the same directory.
-
-SentenceTransformer uses `encode_query`, `encode_document`, and `similarity`;
-CrossEncoder uses `predict` on query/document pairs; MultiVectorEncoder uses
-query/document encoding followed by native MaxSim. Saved retrieval prompts are
-applied, with model-card task/prompt overrides recorded in each script where
-needed. All three paths aim to produce four scores on `neuron:0`, with batch size 4:
-
-```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.",
-]
-```
-
-`PASS` requires native Neuron execution, four finite scores, Mars ranked first,
-CPU score agreement at the same precision (`atol=rtol=0.001`), and a stable repeat
-with the same tolerance. `FAIL(output)` means execution completed but a ranking,
-parity, or repeat check failed. A ranking failure can also occur on CPU; it does
-not by itself indicate a Neuron defect. Loading and forward failures include
-CPU-side incompatibilities, unsupported checkpoint formats, and native-device
-errors; the recorded stage and traceback distinguish them. Scores are in original
-document order and are not comparable between models. Results describe this
-configuration and example, not universal model support or retrieval quality.
-
-The sweep uses native eager PyTorch, without an explicit `torch.compile` wrapper,
-on one logical Neuron device per model on Trainium2. Four independent workers
-use separate physical devices; this is not tensor parallelism. SDPA is preferred,
-with eager attention retried where a checkpoint does not support it. FP32 is the
-default, with BF16 for large models or after allocation failure. Final precision,
-attention, configuration overrides, and earlier attempts are retained. Some
-checkpoints require `trust_remote_code=True`; checkpoint revisions are pinned,
-but remote code referenced from another repository is not separately pinned.
-Transformers is 5.22.1, and Sentence Transformers is pinned to commit
-`6be0abf38d00d70bb3c732131df59ad073ff3768` (6.3.0.dev0). The compiler is
-`neuronx-cc 2.28.4568.0+0163fa9e`; each row and script record the tested stack.
-
-This validates one text retrieval example, not throughput, training, image/audio
-inputs, tensor parallelism, or multi-device execution. The memory columns report
-the highest observed sum of Neuron runtime HBM categories (tensors, code,
-constants, scratchpad, miscellaneous runtime allocations, and DMA rings), sampled
-in-process at a target interval of 100 ms. Monitoring begins before model transfer
-and ends after the final inference is synchronized. Capacity comes from
-`torch.neuron.get_device_properties(0).total_memory`: 24 GiB on this instance.
-Values use GiB (bytes / 2**30), matching existing dataset entries despite the
-Space's `Memory (GB)` label. Short transients between samples may be missed.
-Allocator high-water marks are retained separately in raw results. The installed
-`neuron-monitor` did not discover these native processes, so runtime HBM counters
-were queried directly. Failed runs without a completed capture leave memory
-empty rather than recording zero. Per-category samples are retained under the
-sweep's `memory_samples/` directory.
-
-Wall time covers imports, loading, the CPU reference, first Neuron execution and
-compilation, the repeat, and monitoring overhead. It is not per-query inference
-latency, and concurrent workers can affect timing. `runner` names the encoder
-class; `neuron_cores=1` means one logical device, not a tensor-parallel degree.
-
-Full scores, configuration, prior attempts, and per-view counts are in
-`validation/sentence-transformers/2026-11-03-top100/results.json`, with a compact
-`results.csv` alongside it. Each tested row includes its standalone `run_py`,
-which writes `result.json` and exits nonzero when validation fails. Failure logs
-are retained in `logs/`. No compiled cache was published. The earlier top-10
-study and supplementary MultiVectorEncoder remain available in
-`validation/sentence-transformers/2026-11-03/` as historical evidence.
-
 ## 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 sha256:7e855481a7e894011298c98e89b1cdc5e2be2770f8f07e3e3cd637be7e0db904
-size 882237
+oid sha256:1d81e7876183b15b0a5b0be8a7a0a1fe7fbc3db54df4e912061420f58ba15ed5
+size 884470