Skip to content

Commit 99e4432

Browse files
committed
Record exact but slower prepared-row experiment
Agent-assisted-by: Codex (OpenAI) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>
1 parent eec001d commit 99e4432

4 files changed

Lines changed: 439 additions & 2 deletions

File tree

‎evidence/h3-image-ascend/README.md‎

Lines changed: 11 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -204,3 +204,14 @@ the NPU worker grace through the outer stage client, and adds the explicit
204204
internal-format control. [Checks and limitations](results/review-followup-checks.json)
205205
record 26 CPU regressions and the single-device graph-output lifetime test.
206206
These checks do not requalify historical HTTP measurements on the new revision.
207+
208+
## Round twelve: prepared row indices rejected
209+
210+
Preparing local embedding-row indices once per request preserved exact RGB
211+
output in all 30 matched images. It did not improve HTTP latency: control
212+
W1/W2/W3 medians were 11.039/2.574/4.243 seconds, versus candidate
213+
11.197/2.925/4.502 seconds. [Samples and all quality checks](results/prepared-rows-round12-rejection.json)
214+
record the unchanged workload and source identity. The candidate was not
215+
adopted, and no final ABBA qualification was run. This experiment uses the
216+
older ad37f7f engine base; it does not requalify the 7f8395fed follow-up.
217+
Campaign acceptance remains **INCONCLUSIVE**, with all latency targets unmet.

‎evidence/h3-image-ascend/export-manifest.json‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,5 @@
11
{
2-
"README.md": "fc12834138c87c16ddf9f7204166db579d444fc6735daa953873247cbc3664ee",
2+
"README.md": "da3a669573f9e343e1913cad1c7b232bf9a7921f692b5e60749230436da4196e",
33
"patch/revision.json": "d8396389afd8c9c5565f9834e1a4ab092813e74c78285c921f7453b27a84235b",
44
"plan/environment.json": "9b404a74007446fbbd6343383ebf8d18be446068d2018eb330c964473f54c45e",
55
"plan/provenance.json": "a58206b9cb0f2dc1f0a51c71ce929630a3552a13bd7fdd5ced290a71945c28d6",
@@ -36,6 +36,7 @@
3636
"results/mm-alltoall-real-rank0.json": "8a0eb090272def70c860477d9867a51cf7dbb34e4638952719b2ae05c2177ccb",
3737
"results/native-sp4-offload/request-measurements.json": "586eddd05b217320d86ff8a3d98b0067e1904f3a706e24cf3a2051d8a9da11e2",
3838
"results/native-sp4-offload/summary.json": "5eb77f9bada84dba34b2d097dda158f0e37f5bb9aa92fa776072491b828aa08c",
39+
"results/prepared-rows-round12-rejection.json": "5445aacc75663e99dbbe8193c3dc53491835150018c49db120e122e06f608ac4",
3940
"results/quality-overview.png": "163b7d403cb01ec40a56469f859c7f90c0e9621abddf22f8c740c4baa47ea8d9",
4041
"results/quant-projection-sensitivity.json": "b3ae5166c955bda3a0e6d9b075376e2a31b02c2db315cadaef6bc166ce42c5be",
4142
"results/quant-sensitivity.json": "45de8c82872edd70e0d551741402a676529cc0fc28948cb856c809a2fc9f825a",

0 commit comments

Comments
 (0)