dewml/dory-1.0-4b-vllm-reference

Dory 1.0 4B as vLLM computes it

Reference statistics for the release test of Dew (tests/test_dory.py), recorded by tools/dory_vllm_reference.py from vLLM's own engine running nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16: vllm-project/vllm#60149 at shan18/vllm 55076950b6c155f384d6766d63563c3bc20a0e3f, installed by vLLM's Python-only build over https://wheels.vllm.ai/153a17166e5d07bc38184dedeea542d2e8885e55/vllm-0.32.0rc1.dev307%2Bg153a17166-cp38-abi3-manylinux_2_28_x86_64.whl.

Runs (loop count / cache mode / dtype): 1/per_loop/bfloat16, 1/last_loop/bfloat16, 2/per_loop/bfloat16, 2/last_loop/bfloat16, 4/per_loop/bfloat16, 4/last_loop/bfloat16, 2/per_loop/float32, 2/last_loop/float32. Each decodes a sentence greedily for 256 tokens and 64 tokens after 1100 random ids, past the 1024-token sliding window.

• probe.npz: the prompts, each run's drawn tokens, each row's probe columns with vLLM's and the float64 truth's log-probabilities on them, four rows over the whole vocabulary, and vLLM's own model code run in float64 on the same tokens.
• The orders are tools/dory_vllm_reference.py's symmetry, drawn from their seeds: each a permutation of the residual stream's units and of each layer's hidden units, heads and value dimensions, the first batch's first the identity.
• upstream-<run>.npy: every row over the whole vocabulary from vLLM's own model code in float64 on the run's tokens, the truth.
• columns-<run>-<batch>.npy: under each order of a batch, vLLM's rows on the probe columns, [52, rows, 64] float32.
• rms-<run>-vllm.npy: each order's RMS distance of the run's rows from the float64 truth, [batches, 52]; rows-<run>-vllm.npy, each row's, [batches, 52, rows]; sizes-<run>.npy, each truth row's RMS, the row's size. The release test holds Dew's to these, leaf and row.
• absolutes-<run>-vllm-<batch>.npy: every entry's absolute distance from the float64 truth summed over a batch's 52 orders, [rows, vocabulary] float32, for calibration and the report.
• invariance-<loops>.npy: under each order, the model code's float64 rows on the sentence's prompt, their largest distance from the identity's over their largest magnitude, [batches, 52]: the evidence each order is an exact symmetry.

On a last-loop run's rows past the window these files hold the pull request's model code at the engine's precision (bf16, or fp32 with TF32 attention products) instead of the engine, the reference there: the engine's last-loop cache parts from the model code past the window (https://github.com/vllm-project/vllm/pull/60149#issuecomment-6270937659).

Dew's tests/fixtures/hf/dory-1.0-4b/source.json pins this repository's revision and each file's sha256.

These files are the model's outputs and Dew's measurements of them, released under Apache-2.0. The model itself is released under OpenMDW-1.1 (https://openmdw.ai/license/1-1/), which places no restriction on its outputs; no part of its weights or other model materials is in this repository.