fix: stabilize remaining A5 board samples - #1395
Conversation
|
/run a5 Rowexpandsub Rowexpanddiv Cmp Sel Rowexpandmul Sels Partmin Scatter Cmps Movfp |
|
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 板测成功
|
|
/run a5 rowexpandsub rowexpanddiv cmp sel rowexpandmul sels partmin scatter cmps movfp_fixpipe_reuse |
|
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 sels |
|
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 sels |
|
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 板测成功
|
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b395757fea
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".
| return emitOpError("expects 'scaling_tile' element type to be f16, bf16, " | ||
| "f32, or packed i64/ui64 on A5"); |
There was a problem hiding this comment.
Update the A5 negative payload regression
This change makes the exact deqf16_vec/ui64 module in test/lit/pto/fixpipe_frontend_verify_set_quant_vector_bad_payload_type_a5.pto valid, but that test still runs not ptoas --pto-arch=a5 and expects the old error text. The compiler will now succeed, causing the not command—and therefore check-pto—to fail before FileCheck; replace that negative case with a still-invalid payload or convert it into a positive regression.
Useful? React with 👍 / 👎.
| // A5 pto-isa uses packed uint64_t FBUF entries for VDEQF16, while other A5 | ||
| // vector-quant forms expose direct floating-point scaling payloads. | ||
| return isPackedI64 || elemTy.isF16() || elemTy.isBF16() || elemTy.isF32(); |
There was a problem hiding this comment.
Restrict packed A5 payloads to VDEQF16
When an A5 pipe uses req8_vec, qf322b8_pre_vec, or qs322bf16_pre_vec, this unconditional isPackedI64 branch also accepts an integer payload, even though the adjacent comment identifies packed FBUF entries as the VDEQF16 representation and says the other A5 forms use direct floating-point payloads. Because the caller already resolved quant but does not pass it here, such ill-typed IR passes verification and is forwarded unchanged to SET_QUANT_VECTOR; make the accepted element family depend on the referenced quant mode.
Useful? React with 👍 / 👎.
|
/run a5 sels |
|
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 板测成功
|
4687426 to
7e2ec3e
Compare
修复最近 A5 Nightly Board 中可复现的普通样例失败,保持 DeepSeek/Qwen 相关样例不变。
已处理:
未处理:DeepSeek、Qwen、TquantMx 相关样例。
验证:
python3 -m py_compilegit diff --check通过/run a5 Rowexpandsub Rowexpanddiv Cmp Sel Rowexpandmul Sels Partmin Scatter Cmps Movfp