use f32 for f16 in npy - #2590
Merged
Merged
Conversation
Collaborator
Author
|
/ci full |
Collaborator
Author
|
/ci examples |
Collaborator
Author
|
/ci llm |
✅ CI / full: success
|
kali
force-pushed
the
feat/cli-npz-f16-npyz
branch
from
August 5, 2026 08:40
13d8fd4 to
6aeb0b7
Compare
|
Reference: 2026-08-05 morning nightly run (0d old) · full report → run Speed — evaltime · prefill · decode no inference-speed regressions
|
| Δ | metric | device | main → PR |
|---|---|---|---|
| arm_ml_kws_cnn_m load · pass |
cortex-a9 |
84 ms → 106 ms | |
| arm_ml_kws_cnn_m load+optimize · pass |
cortex-a9 |
130 ms → 155 ms |
kali
changed the title
✅ CI / examples: success
|
|
🔴 Bench vs main — 1 speed regression(s) · Reference: 2026-08-05 morning nightly run (0d old) · full report → run Speed — evaltime · prefill · decode
|
| Δ | metric | device | main → PR |
|---|---|---|---|
| arm_ml_kws_cnn_m load · pass |
cortex-a9 |
84 ms → 119 ms | |
| arm_ml_kws_cnn_m load+optimize · pass |
cortex-a9 |
130 ms → 174 ms | |
| hey_snips_v1 load · 400ms |
cortex-a9 |
77 ms → 93 ms | |
| hey_snips_v1 load+optimize · 400ms |
cortex-a9 |
147 ms → 172 ms | |
| arm_ml_kws_cnn_m load · pass |
cortex-a53 |
47 ms → 54 ms | |
| arm_ml_kws_cnn_m load+optimize · pass |
cortex-a53 |
73 ms → 82 ms | |
| hey_snips_v4_model17 load+optimize · 2sec |
cortex-a7 |
5.12 s → 5.69 s | |
| mobilenet_v1_1 RSS @ ready · pass |
cortex-a53 |
48.1 MB → 53.2 MB | |
| hey_snips_v4_model17 load · 2sec |
cortex-a7 |
4.39 s → 4.83 s | |
| mobilenet_v1_1 load · pass |
cortex-a9 |
2.09 s → 2.26 s | |
| mobilenet_v1_1 load+optimize · pass |
cortex-a9 |
2.37 s → 2.56 s |
✅ CI / large-models: success
|
Collaborator
Author
|
/ci full |
Collaborator
Author
|
/ci llm |
The io.npz bundles held f16 arrays, which the CLI cannot read (ndarray-npy has no f16 support and the writer stores f16 as f32). Up-cast the bundles to f32 and pass --allow-float-casts so the f16 graph boundary reconciles.
Collaborator
Author
|
|
kali
force-pushed
the
feat/cli-npz-f16-npyz
branch
from
August 6, 2026 12:40
6aeb0b7 to
b4b36ae
Compare
Collaborator
Author
|
/ci full |
|
🔴 Bench vs main — 1 speed regression(s) · Reference: 2026-08-06 morning nightly run (0d old) · full report → run Speed — evaltime · prefill · decode
|
| Δ | metric | device | main → PR |
|---|---|---|---|
| hey_snips_v31 load · 400ms |
cortex-a9 |
298 ms → 353 ms | |
| hey_snips_v31 load+optimize · 400ms |
cortex-a9 |
343 ms → 406 ms | |
| arm_ml_kws_cnn_m load · pass |
cortex-a9 |
83 ms → 97 ms | |
| hey_snips_v4_model17 load · 2sec |
cortex-a9 |
3.81 s → 4.25 s | |
| arm_ml_kws_cnn_m load+optimize · pass |
cortex-a9 |
129 ms → 143 ms | |
| hey_snips_v4_model17 load · 2sec |
cortex-a7 |
4.31 s → 4.78 s | |
| hey_snips_v31 load · 400ms |
cortex-a7 |
334 ms → 369 ms | |
| hey_snips_v4_model17 load+optimize · 2sec |
cortex-a9 |
4.39 s → 4.83 s | |
| hey_snips_v31 load+optimize · 400ms |
cortex-a7 |
381 ms → 419 ms | |
| hey_snips_v4_model17_nnef load · pulse8 |
cortex-a55 |
382 ms → 418 ms | |
| hey_snips_v4_model17 load+optimize · 2sec |
cortex-a7 |
5.04 s → 5.51 s | |
| en_tdnn_15M_nnef load · pulse_240ms |
cortex-a53 |
452 ms → 485 ms | |
| hey_snips_v4_model17_nnef load · pulse8 |
cortex-a53 |
471 ms → 502 ms |
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.
ndarray-npy cannot decode float16 arrays, so npz bundles holding f16 (and the writer stored f16 as f32). Switch the CLI npz reader and writer to npyz, which supports every numpy scalar dtype tract has a Datum for, so f16 bundles round trip natively.