test: sync pypto decode samples - #799
Conversation
There was a problem hiding this comment.
Code Review
This pull request updates the Qwen3 decode PTO kernels for both A3 and A5 targets. The generic @qwen3_decode_incore_* function names have been remapped to descriptive kernel names (such as @q_proj, @gate_proj, and @softmax) to align with the current pypto-lib/main raw kernel names. Additionally, minor adjustments were made to tile buffer allocation addresses and constants in qwen3_decode_incore_5.pto and rope_kv_cache.pto, and the README files were updated accordingly. There are no review comments provided, and I have no additional feedback on these changes.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
Codex Review该评论由 review 机器人自动更新。
SummaryReview failed at stage Findings未生成结构化 findings,因为 review 过程提前失败。 Log Tail |
|
/review |
Manual Codex Review该评论由
Summary收到 FindingsReview in progress. |
Manual Codex Review该评论由
SummaryReview failed at stage Findings未生成结构化 findings,因为 review 过程提前失败。 Log Tail |
|
/review |
Manual Codex Review该评论由
Summary收到 FindingsReview in progress. |
Manual Codex Review该评论由
SummaryReview failed at stage Findings未生成结构化 findings,因为 review 过程提前失败。 Log Tail |
60b7e87 to
e52cc94
Compare
|
/run a3 |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e52cc94c22
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| } | ||
| for testcase in DEEPSEEK_V4_DIRECT_CASES | ||
| } | ||
| CASE_INT_SCALAR_DEFAULTS = {} |
There was a problem hiding this comment.
Restore the DeepSeek direct-case set
Replacing the old DeepSeek case table with only CASE_INT_SCALAR_DEFAULTS = {} leaves CASE_POINTER_COUNT_MINIMUMS still expanding DEEPSEEK_V4_DIRECT_CASES later in the module, so the generator raises NameError during startup before it can even parse arguments (python3 test/npu_validation/scripts/generate_testcase.py --help fails). Since run_remote_npu_validation.sh invokes this generator for every validation case, this breaks the whole remote validation flow rather than just the removed DeepSeek samples.
Useful? React with 👍 / 👎.
|
已接收
页面会自动刷新,可以直接看当前阶段、排队情况和最近结果。 |
A3 板测失败
失败用例
|
A3 板测失败详情:PR #799out_proj_residual
down_proj_residual
rope_kv_cache
aiv_merge_norm
|
|
/run a3 |
|
已接收
页面会自动刷新,可以直接看当前阶段、排队情况和最近结果。 |
A3 板测失败
失败用例
|
A3 板测失败详情:PR #799out_proj_residual
down_proj_residual
rope_kv_cache
|
0927285 to
bb40fe3
Compare
|
/run a3 qwen deepseek |
A5 板测失败详情:PR #799aiv_qr_hadamard_quant (4/5)
|
A5 板测失败详情:PR #799aiv_qr_hadamard_quant (5/5)
aiv_kv_rms_norm_rope
|
A5 板测失败详情:PR #799aiv_kv_and_cache_write (1/5)
|
A5 板测失败详情:PR #799aiv_kv_and_cache_write (2/5)
|
A5 板测失败详情:PR #799aiv_kv_and_cache_write (3/5)
|
A5 板测失败详情:PR #799aiv_kv_and_cache_write (4/5)
|
A5 板测失败详情:PR #799aiv_kv_and_cache_write (5/5)
|
A5 板测失败详情:PR #799aiv_exp_h_q (1/5)
|
A5 板测失败详情:PR #799aiv_exp_h_q (2/5)
|
A5 板测失败详情:PR #799aiv_exp_h_q (3/5)
|
A5 板测失败详情:PR #799aiv_exp_h_q (4/5)
|
A5 板测失败详情:PR #799aiv_exp_h_q (5/5)
|
|
/run a3 deepseek qwen |
|
已接收
页面会自动刷新,可以直接看当前阶段、排队情况和最近结果。 |
|
/run a5 deepseek qwen |
A3 板测失败
日志尾部 |
|
/run a3 deepseek qwen |
|
已接收
页面会自动刷新,可以直接看当前阶段、排队情况和最近结果。 |
A3 板测失败
日志尾部 |
|
已接收
页面会自动刷新,可以直接看当前阶段、排队情况和最近结果。 |
A5 板测失败
日志尾部 |
|
/run a5 deepseek qwen |
|
已接收
页面会自动刷新,可以直接看当前阶段、排队情况和最近结果。 |
A5 板测失败
失败用例
|
A5 板测失败详情:PR #799swa_gather_kv
|
Summary
PTOAS/main.ptooutputs directly, preserving the emittedkernels/hierarchy, filenames, and file contents/run a3 deepseek qwenand/run a5 deepseek qwenselect both model groupsProvenance and inventory
hw-native-sys/pypto-lib:008a709dc9a57d11930390c12b1921e371e42ebbhw-native-sys/pypto:93d789bf1727343efc48949581642bee47a683f9628dcf68d3776e61341d09088063e426cd5ef69corccbdc4fa5cafd1eda7784c9585f9dc876791778bexports described by an earlier PR revisionThe 9 DeepSeek families not vendored on either platform are fully accounted for in each DeepSeek README:
CommRemoteOffset_*helpers have not been inlinedcsa_slots_build_valid_qk_planfamily exceeds the 60-second per-case compilation budgettopk,route_sort, andlm_head_greedy_sample) contain mask-patternpto.tgatherwithout the requiredaxisattribute in the direct exportNo vendored
.ptofile was manually repaired to bypass these verifier failures.Validation semantics
OKmeans the NPU output matched an independently generated expected outputDETERMINISM_ONLYmeans no independent golden was available; the first NPU output was captured and a repeated run matched itkv_hadamardandrms_normnow have independent CPU oracles on both A3 and A5DETERMINISM_ONLY, never as correctnessOK92 OK/93 OKtotals are intentionally removed because they mixed these two validation classesValidation performed
--pto-level=level3kv_hadamard: complete independent golden generatedrms_norm: complete independent golden generatedkv_hadamard:OKrms_norm:OKkv_touch:DETERMINISM_ONLYOK/ 1DETERMINISM_ONLY/ 0FAIL/dev/davinci0bash -nfor the changed shell scriptspython3 -m py_compilefor the changed Python helpers and golden scriptsgit diff --check