Repository navigation
[executorch][native-ptn] Write byte-identical PTN packages for identical inputs - #25380
digantdesai wants to merge 1 commit into
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/25380
Note: Links to docs will display an error until the docs builds have been completed. ⏳ 1 Pending, 1 Unrelated FailureAs of commit f444be7 with merge base 90bd617 ( BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR needs a
|
SS-JIA
left a comment
There was a problem hiding this comment.
Review automatically exported from Phabricator review in Meta.
JakeStevens
left a comment
There was a problem hiding this comment.
Diff landed internally; approving.
SS-JIA
left a comment
There was a problem hiding this comment.
Diff landed internally; approving.
…uts to F32 Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0) (oldest at bottom): * #25381 * #25380 * #25379 * #25378 * #25377 * #25376 * #25375 * #25374 * #25373 * #25372 * #25371 * #25370 * #25369 * __->__ #25368 layer_norm widens F16 and BF16 inputs, weight and bias to F32 and normalizes in F32 as eager does; set() rounds the result back. hardtanh (and relu6) widens BF16 to F32, since ggml-cpu clamps only F32 and F16; F16 stays one native ggml_clamp, as in llama.cpp. Rounding half node results at set() left these lowerings refusing (layer_norm) or aborting (BF16 hardtanh) on half inputs that F32-computed lowerings feed. Authored with assistance from Claude and Codex. Differential Revision: D133480458 Pull Request resolved: #25368
…cases Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0) (oldest at bottom): * #25381 * #25380 * #25379 * #25378 * #25377 * #25376 * #25375 * #25374 * #25373 * #25372 * #25371 * #25370 * __->__ #25369 * #25368 test_ggml_ops_cuda and test_ggml_ops_hip run as bundles of 40 cases, so their 165 cases take about 5 test actions instead of one action per case. The CPU variant is unchanged. Each unbundled case paid ~19 s of torch/transformers import for ~3 s of test, so the GPU suites ran past their CI time limit. Nothing hangs or fails. Authored with assistance from Claude and Codex. Differential Revision: D133480459 Pull Request resolved: #25369
Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0) (oldest at bottom): * #25381 * #25380 * #25379 * #25378 * #25377 * #25376 * #25375 * #25374 * #25373 * #25372 * #25371 * __->__ #25370 * #25369 * #25368 run_hf_llm_ggml and ggml_ptn_runner take their engine flags (--device, --device_index, --cpu_threads, --max_cached_graphs, --print_graph_ops) from one GgmlEngineFlags library, and the README gains a table of the engine options that change numerics. Both runners then accept the same flags with the same meaning and help, and a later option is added and documented once. Authored with assistance from Claude and Codex. Differential Revision: D133480460 Pull Request resolved: #25370
Stack from ghstack (oldest at bottom):
write_ptnandwrite_samplesfix every member's zip metadata (earliest ziptimestamp, owner read/write, Unix host); the package layout and the PTN schema
are unchanged.
Two exports of the same model gave .ptn files that differed only in zip member
timestamps, so a file hash could not show that an export is deterministic.
Authored with assistance from Claude and Codex.
Differential Revision: D131566108