add dynamic_mx_quant test kernel on tag ops-20260828

[ziyang-cheng]

No description provided.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c58b141b00

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Declare QEMU global before using it

Move global QEMU to the beginning of main() (or avoid mutating the module variable). Because QEMU is referenced as the default for parser.add_argument() earlier in the same function, Python rejects this entire script at parse time with SyntaxError: name 'QEMU' is used prior to global declaration, so no precision-check configuration can run.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Pass the compiler environment to make

Pass this env mapping into the subprocess invocation. When a caller supplies the required --compiler-dir argument without separately exporting COMPILER_DIR, the local assignment is discarded because run() invokes subprocess.run() with the inherited process environment; Makefile.common then stops with COMPILER_DIR is not set, making the documented CLI argument ineffective.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Propagate precision failures through the process status

Return a nonzero status when any configuration errors or reports a failed comparison. Currently exceptions are converted into summary strings, comparison output is never interpreted, and main() falls through successfully, so automation receives exit code 0 even when QEMU failed, files were missing, or numerical comparison printed fail; this can allow broken experiment results to pass validation.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Skip compile-only kernel calls during result checking

Guard these throwaway instantiations with #ifndef RES_CHECK, as the non-tail cuBLAS driver already does. With res_check=on, the harness currently executes the half and fp32 calls described as compile-only—and tail_ocp_fp4.cpp and nontail_ocp_fp4.cpp do the same, including the large-block route—before writing the primary result files; in emulator/toolchain combinations where one of those explicitly runtime-blocked branches aborts, an otherwise valid primary kernel can never complete its precision check.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Select a TileN compatible with the OCP Post dimension

Choose a legal aligned divisor of Post when dispatching to the non-tail OCP-FP4 path, or add a partial-column path there. For example, with bf16, BlockSize=32, and Post=192, this helper returns TileN=128, after which nontail_ocp_fp4_plain fails its Post % TileN == 0 assertion even though TileN=64 satisfies the alignment, tile budget, and dimension constraints; valid OCP shapes are therefore rejected solely by the tile-selection heuristic.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Mask the wide OCP maximum before narrowing to bf16

For half or fp32 input, clear the fp32 exponent/mantissa bits before converting the reduced maximum to bf16, as the tail OCP implementation already does. If a block maximum lies just below a power of two—for example the fp16 value 1.9990234375—this TCVT rounds it up to 2.0 before the subsequent bf16 mask, whereas the required exponent extraction and the golden generator truncate the original value; the resulting shared scale is one exponent too large and the entire block is quantized at half the intended magnitude. The same ordering appears in the big-BlockSize OCP path.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Require exact equality for compact E8M0 scales

Use exact byte equality (or require max_ae == 0) for the compact scale result. This function documents that the compact E8M0 layout is directly byte-comparable, but the mse < 0.1 threshold lets sparse errors pass—for example, one off-by-one scale byte in a 64-byte output has MSE 0.015625 and is reported as pass—so incorrect block scales can escape precision validation.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Build the supported tail OCP-FP8 kernel in compile.all

Add TYPE=TAIL_OCP_FP8 to this build list. The repository-wide compile_all.sh invokes this compile.all, the scoped Makefile defines a TAIL_OCP_FP8 target, and the accompanying README lists that production kernel among the five debugged implementations, but this script still builds only the older set of four; consequently the normal full-compilation check never compiles the newly supported kernel and cannot detect regressions in it.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 71abf61567

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Match the TAIL_OCP_FP4 generator dimensions

When TAIL_OCP_FP4 is selected, this entry generates an 8×64 input and golden output, but tail_ocp_fp4.cpp is compiled without PM/PN overrides and therefore processes its 512×256 defaults. Its readBinaryFile(..., sizeof(xbits)) rejects the short generated input, while the driver ignores that failure and writes 512×256-sized result files, so the comparison can only report size mismatches rather than testing this kernel. Either configure this entry as 512×256 or pass matching PM and PN definitions during compilation.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Treat matching FP8 NaNs as equal

When an FP8 golden result and the kernel output both contain the E4M3 NaN encoding (0x7f or 0xff), both bytes decode to NaN, so their subtraction makes mse and max_ae NaN and this comparison reports fail even though the outputs match. This prevents the comparator from validating the special-value paths that the generator and kernels explicitly implement; compare matching NaN lanes by encoding or mask them as equal before computing the metrics.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Reject incomplete reduction blocks

If the tail reduction dimension (cols) is not divisible by block_size, this floor division silently omits the final partial block; the nontail branch does the same for rows. compute_golden nevertheless emits full-sized files with every omitted element left at its initialized zero value, producing a plausible but incorrect golden result instead of reporting that the shape violates the kernels' exact-block contract.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Reject unsupported algorithm and dtype combinations

The CLI accepts the algorithm and destination dtype independently even though the module documents only OCP with FP8/FP4, CUBLAS with FP8, and DYNAMIC_RANGE with FP4. Invocations such as --algo CUBLAS --dtype FP4 or --algo DYNAMIC_RANGE --dtype FP8 therefore complete successfully and write golden files by combining unrelated scaling and encoding rules, despite there being no corresponding operator semantics. Validate the pair after parsing so unsupported experiments fail rather than consuming fabricated reference data.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Store tail-column scales at the full-tile offset

For a supported nontail cuBLAS shape with both full and partial column tiles, such as Post=96 and derived TileN=64, this tail scale iterator has physical width N_tail=32, so s_iter_r(0, numN) advances by numN * N_tail and writes at column 32 rather than the required numN * TileN column 64. The tail scales overwrite part of the preceding full tile and the actual trailing columns remain unwritten; fold numN * TileN into the base pointer and use column index zero, as the other folded scale stores do.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Initialize the compact scale padding

Whenever numKb is odd, including the default one-block case, scaleCols reserves an extra padding byte but the kernel stores only columns 0..numKb-1. The test drivers' statically zero-initialized arrays hide this, while a normal caller supplying reused or uninitialized output memory receives a stale padding byte instead of the required compact-layout zero. Explicitly store zero into the padded column, as dynamic_mx_quant_tail_ocp_fp4 already does; the same omission should also be corrected in the other implementations that describe their padding as merely "left zero."

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 150660a5dc

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Force res-check builds to recompile their objects

These commands run immediately after plain builds of the same sources, but changing res_check only changes compiler defines and is not represented in the object-file dependency graph. The clean recipe searches for dynamic_mx_quant.o rather than objects such as tail_cublas_fp8.o, so these invocations relink the plain objects without RES_CHECK/ENABLE_BINARY_OUTPUT; the resulting ELFs do not read inputs or write result files. Remove the relevant objects or use configuration-specific object paths before every res-check build.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Preserve representable E4M3 subnormals

The E4M3 minimum positive subnormal is 2^-9, so round-to-nearest should only produce zero below the midpoint 2^-10. For example, 0.0015 is closer to 2^-9 than to zero, but this branch emits 0x00; precision data containing sufficiently small scaled values therefore gets an incorrect golden result. The following subnormal calculation also needs to quantize in 2^-9 increments rather than 2^-6 increments.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Reject individual FP4 output mismatches

Using only aggregate MSE lets incorrect packed FP4 results pass. In the standard 8×64 case, one nibble differing by the minimum nonzero magnitude step of 0.5 contributes only 0.25/512 MSE, far below 0.1, while the computed max_ae is ignored; consequently a kernel can corrupt isolated outputs and still be reported as passing. Require exact codes or enforce an appropriate per-element/MaxAE bound in addition to MSE.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Synchronize 4-PE result writes

When this driver is run with its required multiThreadNum=4, every PE executes main() and reaches these calls after computing only its own row partition. All four threads therefore open and truncate the same files concurrently, and an early PE can read the shared buffers for write() while other PEs are still updating them, yielding partial or interleaved precision results. Synchronize completion and restrict file output to one designated I/O thread, as the existing multi-thread test drivers do.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 65bc49e848

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Propagate failures from every dynamic-mx build

Make this script preserve a nonzero status from any failed make invocation. As written, its exit status is only that of the final FP4 probe, so if any production or res_check target on lines 12–21 fails while that probe succeeds, the newly wired benchmark/one-level-arch/compile_all.sh treats the operator as successfully compiled and the full-compilation check misses the regression.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Honor kPeNum when routing to the big-block kernel

When kPeNum=4 and the shape selects the big-block route (for example cuBLAS FP8 with BlockSize=128), this call drops kPeNum; the big-block implementation has neither get_thread_idx() nor a partition and loops over every kb. In a four-thread SPMD launch, all four PEs therefore recompute the entire tensor and concurrently store to the same y and scale addresses instead of dividing the work. Partition the big-block loop, restrict it to one PE, or reject this combination; the OCP-FP4 dispatcher has the same omission.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: cd71c3cbfa

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Remove stale outputs before each precision run

Clear output.bin and scale_output.bin when preparing this directory, including when --skip-gen is used. On a repeated run, if QEMU exits successfully without rewriting either file—for example because a driver ignores a failed writeBinaryFile call—compare_results() reads the previous run's files and can report a pass for a run that produced no results.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Reject isolated FP8 output mismatches

Require exact FP8 codes or enforce a per-element/MaxAE bound in addition to aggregate MSE. For the standard 8×32 case, changing one expected 1.0 value to 1.125 contributes only 0.125²/256 ≈ 0.000061 MSE, so this condition reports pass despite a wrong FP8 output byte; localized kernel corruptions therefore escape the precision check.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Infer FP4 for suffixed driver names

Recognize FP4 anywhere in the driver suffix (or make dtype optional before inference). When the comparator is invoked directly without --dtype for the introduced dynamic_mx_quant_nontail_ocp_fp4_4pe.elf, dtype defaults to FP8 and the basename does not end with fp4, so packed FP4 bytes are decoded as E4M3 values and the reported metrics/status do not validate the kernel's actual output.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1b8c745268

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Launch the four-PE precision case with four threads

The registered TAIL_OCP_FP4 check is launched with this single-PE QEMU argument set, but its kernel hardcodes kPeNum = 4 and assigns each PE a separate row segment; with only PE0 running, three quarters of the output and scale remain unwritten. The precision runner therefore cannot exercise this configuration correctly unless it supplies the documented four-thread setting or uses a single-PE kernel variant.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Restrict the FP4 kernel to packed output types

If a caller overrides the public OutT template with a supported non-FP4 type such as __fp8_e4m3, recip_emax_bits<OutT>() accepts it and the kernel compiles, but this global layout and the stores below still use N / 2 and BlockSize / 2 byte offsets. FP8 output then overlaps blocks and rows instead of occupying one byte per element; add an assertion limiting OutT to the two packed FP4 types or derive the storage layout from its width.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Reject filters that select no configurations

When the requested filters have no intersection—most directly with the accepted --algo DYNAMIC_RANGE, for which CONFIGS contains no entry—this comprehension produces an empty list, the test loop is skipped, and the script prints an empty summary before exiting successfully. A mistyped or currently unsupported experiment can therefore appear to pass without compiling or running anything; reject an empty types_to_run selection with a nonzero status.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Propagate NaNs while reducing cuBLAS golden blocks

When a block contains a NaN after a finite value, Python's order-dependent max can retain the finite value—for example, max([1.0, nan]) is 1.0. The golden generator then emits a normal scale and reciprocal instead of the cuBLAS path's required 0xff scale and NaN reciprocal, so special-value tests depend on the NaN's position and can disagree with the bit-domain reduction used by the large-block kernel; detect any NaN explicitly before reducing the finite magnitudes.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Point the comparator default at the generated compare tree

When the comparator is invoked directly without --cmp-root, this default resolves from .../dynamic_mx_quant/src to benchmark/one-level-arch/test/kernel/compare, while both the runner and generated artifacts use benchmark/one-level-arch/compare. The documented standalone -d/-l usage therefore cannot find its output and golden files unless callers manually override the supposedly optional path; derive the default from the one-level benchmark root instead.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 837d2c5594

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Serialize dynamic four-PE result dumping

When this driver runs with its default PPE=4, all four PEs execute main() and reach these calls after computing only their own row partition. Each PE therefore opens and truncates the same two files while other PEs may still be updating the shared buffers, producing partial or overwritten precision results. Use the SpmdResCheck<4> leader-only I/O protocol already used by the static four-PE drivers.

Useful? React with 👍 / 👎.