campaign.sh fans a screened operator list across cards, one ab_repeat.sh per
card, serial within a card. It does no measuring: ab_repeat.sh is the path that
produced the four operator verdicts already on record, and it stays byte for
byte as it was.
Each card gets its own AB_ROOT. Eight ab_repeat.sh started in the same second
would otherwise compute the same timestamped directory name and two cards would
write into one.
ZL_ISOLATE_TUNER_CACHE=1 is forced, since eight processes on one sqlite is the
thing to avoid. The cost is that pinned configs differ from a shared-cache
batch, so absolute latencies cannot be quoted across the two; the caliber guard
refuses such comparisons on its own, which is why this is a note and not a
warning someone has to remember.
partition.py's shell output gained the per-card weight, because EXTRA_OPS has
to land on the lightest card and counting operators gets that wrong — one heavy
operator outweighs three light ones. It writes through stdout.buffer: Python
turns "\n" into CRLF on Windows, and the trailing \r then rides into the last
field and breaks the numeric comparison that picks the card.
Two pieces the parallel campaign needs.
partition.py assigns screened operators to cards longest-first. run_batch.py's
existing hash partition ignores weight: on this batch it puts the critical path
at 120 minutes and leaves a card idle, where greedy gets 70. The floor is set
by the single heaviest operator — topk alone is 70.6 minutes against an ideal
share of 45.4 — so partitioning cannot beat it, only keep the other cards from
waiting. Weights come from screen.tsv's elapsed_before + elapsed_after, which
is one pair, i.e. 2 cells; N rounds is 2N cells, so the estimate is N times the
weight. Getting that wrong reported topk as 11.8 minutes.
RESCORE=<root> re-judges an existing screening root against the current
fingerprint and rewrites its screen.tsv, keeping the original alongside. The
criterion has already changed twice — index.tsv used to be hashed, and the
switch's own ttg.rlc-* annotations had to be stripped — and the second change
moved 24 operators from 23 IR_CHANGED to 16. Rescoring the artifacts already on
disk takes seconds; re-running that screen takes 79 minutes.
FlagTree's _apply_musa_rlc_policy(mod) stamps the whole RLC policy onto the
module as ttg.rlc-* attributes whenever the switch is on, regardless of whether
the pass then changes anything. Hashing the raw text therefore carried a
constant offset: every operator that compiles a kernel differed on that one
line and screened as IR_CHANGED. A filter that excludes nothing is not
filtering.
The null control could not catch it. Both of its sides run with the switch off,
so both carry the same attributes and the fingerprints match — it proved the
dump is deterministic, which was true, and said nothing about this.
Stripping ttg.rlc-* before hashing turns the screen back into a filter. Over
the same 24 operators, recomputed from the artifacts already on disk: 23
IR_CHANGED becomes 16, with 7 falling back to NO_IR_DIFF — tanh, gelu, mul,
softmax, mm, bmm, replication_pad3d, each byte-identical apart from that line.
replication_pad3d is the PR's largest NVIDIA win at +124.6%; on MUSA this pass
does not touch it.
The counts also show why the convert_layout count fails as a criterion in both
directions: addmm keeps cvt at 15 -> 15 while all 15 of its ttgir files differ
in substance, and tanh has no conversions and no change.
FlagGems keeps per-vendor implementations under
flag_gems/runtime/backend/_<vendor>/ops/, and benchmarks commonly gate on
whether one imports:
try:
from flag_gems.runtime.backend._nvidia.hopper.ops.mm import router_gemm
ROUTER_GEMM_AVAILABLE = True
except Exception:
ROUTER_GEMM_AVAILABLE = False
An import only proves the .py is on disk. It touches no device, so it succeeds
on every backend and the gate never fires. test_router_gemm.py therefore ran
the NVIDIA Hopper kernel on an S5000: its autotune and heuristics are keyed to
nvidia, supply no BLOCK_M/N/K/SPLIT_K on musa, and the run died deep inside the
JIT with "dynamic_func() missing 4 required positional arguments" — an error
that points nowhere near the cause. Meanwhile _mthreads/ops/mm.py holds a
working router_gemm the benchmark never imports. Gating on vendor_name first,
the way the Cambricon branch in the same file already does, makes it run: 10
shapes, 10 kernels, cvt 10 -> 2.
The plugin now scans sys.modules at session end and prints
FOREIGN_VENDOR_MODULE with the vendors it found. It reports rather than
refuses: loading another vendor's module is not proof of a problem, but when a
result looks wrong it is the first line to check. Vendor roots are compared
segment-wise, not by prefix, so a vendor whose name extends ours is not
mistaken for our own.
README gains the ordered checklist for an operator that will not run, with the
real instance behind each entry.
A pair judged incomparable was still followed by a speedup range on the next
line. A NO_TRITON pair's range says nothing about the switch — nothing reached
the compiler — and a NO_IR_DIFF pair's range is the machine's noise floor for
that operator, which is useful but is not a result. Printing either unlabelled
produces exactly the thing this tooling exists to prevent: a table that looks
like a finding. NO_ROWS and NO_TRITON now print no range; NO_IR_DIFF prints one
and says what it is — how far it sits from zero is the confidence floor for a
single measurement of that operator.
screen.tsv gains the seven caliber columns, read back through
`batch_summary.py --caliber`, so CALIBER_FIELDS stays the single definition and
the screening table does not restate it. Two screen.tsv files taken at
different --level are otherwise indistinguishable side by side, and those
columns are the only thing that would stop the comparison.
Screening is first contact with an operator, which makes it the place most
likely to meet one that hangs — and it had no upper bound at all, so a single
operator could hold the whole pass indefinitely. The entire argument for
having a screening step is that it is fast.
CELL_TIMEOUT (default 2400s, as in ab_repeat.sh) now caps each side, and a
truncated pair says so instead of being folded into the dtype explanation: rc
124/137 is a cost problem, not a verdict about the switch.
elapsed_before/elapsed_after join screen.tsv. The two-step design rests on the
ratio between a pair and six rounds — 0.8 minutes against 13-44 on S5000 — so
the pass should record its own half of that ratio rather than leave it to be
reconstructed from directory timestamps.
"NO_KERNEL" reads as "this operator has no kernel", which is not what the
verdict means and not usually true. It means the run compiled no Triton
kernel: the operator may have run perfectly well through aten or a vendor
library, and a GPU kernel certainly executed. What is absent is anything for
the Triton compiler pass under test to act on.
The note now says that, and says the one alternative explanation is already
ruled out: run_pytest.sh gives every run its own TRITON_CACHE_DIR, so zero
.ttgir cannot be compilation skipped by a warm cache.
zl_bench already had the screening mechanics — ab_rlc_test.sh runs exactly one
config-pinned pair per operator (before records, after replays it) against the
vendor-prefixed RLC switch. I had started a second script for the same job
instead of extending this one; that is removed.
What the existing script could not tell you is whether the pair is worth
following with six rounds, which costs 13-44 minutes per operator against
about 0.8 for the pair. It now writes a verdict per operator to
<root>/screen.tsv, and three of the four are not "the switch does nothing":
IR_CHANGED the two sides' ttgir differ -> measure it
NO_IR_DIFF byte-identical, configs pinned -> skip it
NO_KERNEL rows but no .ttgir -> never reached Triton here
NO_ROWS no measurements -> usually the dtype
The criterion is the ttgir fingerprint, not the convert_layout count: equal
counts do not mean equal generated code, and what the A/B measures is whether
the generated code changed. Counts are still recorded — they say what kind of
change it was. The fingerprint comes from _ir_fingerprint.sh, shared with
ab_repeat.sh's round-1 check so the two cannot drift apart.
OPS entries take an optional per-operator dtype and default to the operator's
own set. Forcing one dtype across a list is how fft got written off as
unrunnable on MUSA: it only accepts complex64.
Output moves under runs/ab_rlc_<ts>/<op>_{before,after} with OUT_DIR set
explicitly, rather than recovering the directory with `ls -dt runs/<op>_*`.
That lookup is correct while one process runs operators in series, which is
the only way it has been used; it stops being correct the moment several cards
share runs/.
Six rounds of A/B cost 13-44 minutes per operator on S5000 (rand, 15 shapes:
13.0 min; sort, 25 shapes: 43.4 min). Screening one costs about 0.8 minutes —
21 operators in 15.0 minutes, which knocked 13 of them out of the expensive
queue and cut it from ~7 hours to ~2.7. That ratio is the whole argument for
having a separate first step.
Screening was being done by ad-hoc scripts living in the results tree, and
those got two things wrong that this codebase already knows better than:
Both sides ran `record`. Each swept autotune independently, so the two sides
compiled different configs and produced different IR for reasons that have
nothing to do with the switch — every operator would screen as "changed".
screen.sh pins configs exactly as round 1 of ab_repeat does: A records, B
replays A.
The criterion was the convert_layout count. Equal counts do not mean equal
generated code, and what the A/B measures is whether the generated code
changed. The ttgir fingerprint decides; the cvt counts are still recorded,
because they explain what kind of change it was.
Four verdicts, and three of them are not "the switch does nothing":
IR_CHANGED (measure it), NO_IR_DIFF (skip it), NO_KERNEL (the operator never
reached Triton on this backend), NO_ROWS (it did not run — usually the dtype).
The last two are runnability problems being mistaken for switch results, which
is how `fft` got written off: the screen forced DTYPES=float16 on every
operator and fft only accepts complex64. OPS entries now carry an optional
per-operator dtype and default to the operator's own set.
ir_fingerprint moves to _ir_fingerprint.sh so screening and the A/B cannot
drift apart on the criterion.
ir_fingerprint hashed every file under the dump directory, which includes
index.tsv and naming.md. An operator that compiles no Triton kernel at all
leaves only those two, identical on both sides, so the switch-effectiveness
check reported "byte-identical ttgir" and skipped the operator with
SWITCH_NO_IR_DIFF — "the switch changed no generated code" — when the truth
was that nothing was generated. router_gemm on MUSA is exactly this case: its
kernel cannot be instantiated there (missing BLOCK_M/N/K/SPLIT_K from the
autotune path), zero .ttgir files, and the check called it a no-op switch.
Now only *.ttgir files are hashed, and a run with none returns failure so the
caller treats it as "no comparable IR" rather than "identical". The message on
that path says so, and says explicitly that it does not mean the switch is
inert.
Piping `telemetry.py --run` into anything that stops reading first — head, or
the campaign shutting the sampler down — raised BrokenPipeError and printed a
traceback. Under ab_repeat that traceback lands in telemetry.err, where it
reads as "telemetry is broken" rather than "telemetry ended normally", and
telemetry.err is exactly the file someone checks when the clock record looks
short.
Swallowing the exception alone is not enough: the interpreter flushes stdout
again on exit and emits a second "Exception ignored" to stderr, so stdout is
pointed at devnull the way CPython documents.
Both the _timing_plugin docstring and the README env table still described the
behaviour from before MUSA declared supports_cuda_graph: "other backends do
not attempt CUDA Graph at all", "非 CUDA 即使设 1 也拒绝". Capture selection has
been driven by the backend capability, not by "is this CUDA", since the MUSA
fallback policy landed. Reading the old text is most likely why the S5000
campaigns were launched with ZL_ENABLE_CUDAGRAPH=0 without anyone checking
what auto would do.
Also states what timer=cuda_graph does and does not mean, since the two are
easy to conflate: it says capture was attempted. Measured on S5000 with auto
and configs pinned, dropout/rand/masked_select select the graph path and then
fall back on every single shape — BENCHMARK_DIRECT_NO_CUDAGRAPH on all 15
rows, zero poisoned, zero rows lost — so the whole table is direct-timed under
a cuda_graph label. ZL_ENABLE_CUDAGRAPH=0 selects the same path and records it
plainly, which is why the existing campaign data stands as measured.
A 6-round A/B here reports side noise at the 0.01% level. Clock throttling and
thermal drift move latency by far more than that and leave no trace in
run.log, so the report comes out looking exactly as clean as a run on a stable
machine. The NV run that reproduced PR763 sampled nvidia-smi every 5s
alongside the campaign; the MUSA campaigns had no equivalent, and the parallel
plan makes it worse by heating the whole box.
telemetry.py samples through the backend, so cuda and musa emit the same
columns (index, sm_clock_mhz, mem_clock_mhz, temp_c, power_w, util_pct) and
two machines' campaigns can be read side by side. mthreads-gmi offers no CSV
query mode, so the MUSA path parses --json and strips the units glued onto
each value ("36C", "1888MHz", "477.68W"); its "Power Draw " key carries a
trailing space, matched by stripping rather than by literal.
The device comes from the backend's own visibility variable when not pinned. A
list of devices resolves to no -i at all, sampling every one, rather than
silently picking the first. Dropping the placeholder drops its flag too: a
bare trailing -i fails the sampler instead of widening it.
Telemetry is evidence, not a gate. A missing or unpermitted sampler exits
quietly with its reason in telemetry.err and the campaign runs unchanged.
The three markers the README requires be checked — AUTOTUNE_REPLAY_FALLBACK,
BENCHMARK_DIRECT_NO_CUDAGRAPH, BENCHMARK_CAPTURE_POISONED — were only
reachable by grepping run.log afterwards, which happens when someone
remembers to. The NV run that reproduced PR763 wrote a per-cell status row
instead; a 6-round campaign here had no equivalent, so a missing cell, a
stalled cell, or a marker count going from 0 to nonzero was invisible until
the batch ended.
run_cell now writes round/op/side/switch/rc/elapsed_s/success/failed plus the
three marker counts to <root>/status.tsv as each cell completes, and wraps the
cell in CELL_TIMEOUT (default 2400s, matching the NV run) so one hung cell
cannot hold the batch. rc 124/137 is called out by name.
grep -c prints 0 and exits 1 when it finds nothing, so each count swallows its
status; without that the bookkeeping itself would mark a clean cell failed.
triton normally comes from a PYTHONPATH pointing at a compiler build and
flag_gems from a source checkout, so a shell that reaches every correct path
can still be the wrong Python. run_pytest.sh checked that the FlagGems
directory and benchmark file exist and then handed the run to pytest, where
the real problem surfaced per cell as `pytest exit 4` from a conftest
ImportError.
A multi-round campaign burns every cell that way before anyone reads a log:
today's noise-calibration run lost 12 measurement cells and 160 load runs to
`ModuleNotFoundError: No module named 'triton'`, each failing in under a
second and each looking, from the outside, like a cell that ran.
The check names the interpreter it used, since which Python is on PATH is the
whole question. Its stderr redirect stays on the child — putting `2>/dev/null`
on the preflight call swallows preflight's own `>&2` message and turns a named
refusal into a bare exit 3.
Two runs on different calibers produce a report that looks exactly like a
valid one — no step errors, the numbers just mean something else. This repo
has published three figures that were an artifact of such a difference:
router_gemm -8.2% (record vs record), mm "945 variants -> 33" (shared vs
inherited libtuner cache), replication_pad3d +43% vs +124.5% (timer).
batch_summary --ab / --ab-rounds now reads every participating run's
run_identity.json and compares backend/device/timer/bench_level/shape_mode/
tuner_cache/direct_repeat before parsing a single latency. Mismatch, or a
missing identity, prints !!! [caliber] and exits 4. autotune_mode is judged
separately: record pinning a replay is the normal shape, both sides record
is the exact trap above and is refused by name. --allow-caliber-mismatch
releases it, still printing what it found.
Recording on a *shared* tuner cache warns rather than refuses: the pinned
config depends on what else ran on the machine, which leaves relative
conclusions intact and absolute latencies unquotable. Refusing it would
reject the default every A/B here currently uses.
ab_repeat.sh closes the one axis that had no check: after round 1 it
fingerprints both sides' ttgir (relative path + content, so a changed
variant *set* counts too). Byte-identical means the switch changed no
generated code for this op, and the remaining N-1 rounds would measure
noise and report it as "no difference" — so the op is skipped with
SWITCH_NO_IR_DIFF unless SWITCH_ALLOW_NO_IR_DIFF=1.
README leads with a caliber table: each axis, its measured magnitude, what
pins it, and which code checks it. The fourth column is code because
documentation did not hold — the README already said "the B side must use
REPLAY_FROM" when the record-vs-record number was published.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A single pair cannot say whether its own verdict is stable, and on this device
that is not a theoretical gap. Measured on an S5000, topk eliminated 32
convert_layout ops yet 17 of its 42 shapes changed the SIGN of their delta
between two equally valid runs. The cause is not the timer: graph capture and
direct timing agree to 1.00x on that operator, and the verdict moved a median of
1.1pp between them, inside the 1.7pp noise floor a same-timer control
established. The deltas are simply near zero, and one measurement cannot orient
them.
ab_repeat.sh runs an A/B over N rounds in independent processes. Round 1's A
side records the autotune configs and every later run -- including A's own later
rounds -- replays that one set, so configs are pinned across rounds and sides and
only the switch under test and measurement noise remain. Round order alternates
to cancel one-directional drift. An operator with a failed round gets no report,
since a report missing rounds reads exactly like a complete one.
batch_summary --ab-rounds aggregates <root>/r<k>_<A|B>/run.log and separates the
two dispersions, which answer different questions and were previously conflated:
the spread within one side's samples says how noisy the machine was, while the
spread of the per-round deltas says whether the verdict holds. Shapes whose
delta changes sign across rounds are counted and called out, because their
direction is not a conclusion at any timer.
The path is already validated on NVIDIA: six rounds narrowed router_gemm to
-16.6% ~ -16.2% with 0.4% side noise, which is what turned a single pair into a
statement.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
do_bench already returns a median over the iterations inside one call, which
suppresses iteration-level jitter. It cannot see drift across a call: clocks
ramping, another process arriving. Direct timing is where that shows, because
it walks the full Python launch path once per iteration, so on small kernels the
number carries launch overhead and CPU jitter that graph capture removes.
ZL_DIRECT_REPEAT=N repeats the call and takes the median over calls, which is
the layer the inner median misses. It defaults to 1, so nothing changes unless
asked.
The knob covers both ways a row ends up on the direct timer: a run that selected
timer=direct outright, and the rows a captured run falls back on, whose marker
now carries repeat=N since those rows were already a different caliber from the
captured ones and can now also be a different sample count. Repetition is
skipped when the caller asks for quantiles or return_mode "all", because the
result is then a sequence and a median across calls is undefined for it.
On MThreads this is not a refinement. A convert-reduction screen over 21
operators puts the largest reductions at sort (65->15), dropout (48->0) and
rand (24->0), and all three are exactly the operators that cannot be
graph-captured there -- RNG seeding and a capture-illegal stream op. The
headline numbers on that device are produced by the noisiest timer, which until
now had no repetition at all.
_cudagraph_plugin imports triton at module scope, so the tests load it against a
stub: the repeat logic is arithmetic and should stay testable without a GPU.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The worker carried a hand-copied plugin array, taken from zl_bench before
backends/ existed, and it had drifted twice over. It still named
_device_guard_plugin, which is now a shim for _device_setup_plugin. More
consequentially it carried no timing plugin at all, so every measurement this
campaign ever sealed was timer=direct while zl_bench's own driver had been
capturing -- a difference the harness never reported, because the missing plugin
cannot announce itself.
It now calls plugin_plan.py with ZL_BACKEND=musa and uses whatever comes back,
so the caliber follows the backend's declared capabilities instead of a snapshot
of them. The plan is preflighted for non-emptiness, since an empty array would
silently run pytest with no plugins at all: no shape injection, no autotune
record, no IR dump, and a table that still looks well-formed.
Two effects on this campaign. _timing_plugin now enters the list, and since
MUSA declares graph capture the worker will capture where the operator allows
it; operators that cannot are marked per row, as on any other backend. And
_parallel_warmup_plugin joins, inert unless PARALLEL_WARMUP_GPUS is set.
The sealed 32-workload run therefore sits on the far side of a caliber change
and is not comparable with anything produced after this commit. Re-running it is
a campaign decision, not a harness one.
The worker test asserted the literal "-p _musa_benchmark_compat_plugin"; it now
asserts the delegation and that no -p is spelled out locally. What the plan
holds for musa is covered where it belongs, in zl_bench's test_backends.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Brings in campaigns/musa_rlc_product/: the runner, its worker, manifest
validation, the sealed 32-workload manifest, a topology re-accounting tool and
their tests. Developed on a fork branch that never reached main, and it merges
without conflict because it only adds a new directory; the one file it shares
with main, _musa_benchmark_compat_plugin.py, is byte-identical to the copy
already committed here.
The campaign layer sits above zl_bench rather than inside it: freezing a
workload set, classifying outcomes, and producing receipts, where zl_bench turns
one operator into one reproducible measurement. Two of its ideas are worth
naming because they are absent below it. Narrowing what a workload measures --
which correctness variant, which benchmark case -- requires a stated reason that
manifest validation enforces, so an exclusion stays visible in review. And
reaccount_musa_rlc_topology.py re-derives verdicts from sealed artifacts under a
corrected accounting rule without re-running anything, while refusing to promote
a workload on the new count, only to demote it.
Its worker still hardcodes a plugin list copied from a pre-backends zl_bench,
which is why the campaign has been timing with timer=direct. Switching it to
plugin_plan.py is the next change and will alter its caliber, so the sealed run
is not comparable across that boundary.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three pieces from the fork's feature/musa-benchmark-support branch that never
reached main, rebased onto the backend abstraction rather than copied verbatim.
_musa_benchmark_compat_plugin drops FlagGems benchmark skip guards written as
`torch.cuda.is_available()` where the code under test dispatches through
flag_gems.device. On MThreads those turn a working operator into "not measured":
test_fft.py skipped entirely, and unskipped on an S5000 it produces 25 rows. The
plugin pins each override to one file and one exact skip reason and only acts
when flag_gems has dispatched to the mthreads vendor; two guards were probed on
device and deliberately left in place, which is why they are named in its
docstring rather than silently absent.
It is wired through a new DeviceBackend.compat_plugins rather than a name test
in the planner, so the policy stays where every other backend difference lives.
Compat plugins are planned first because they rewrite collection-time markers
and nothing may read the item list before them; a test pins that position and
another pins that they do not leak to other backends.
batch_summary gains --ab: the speedup column compares gems against torch, not
one build against another, so an A/B needs the two run.logs paired by operator,
dtype and size detail. Only rows that succeeded on both sides with a positive
latency count, so an incomplete pairing cannot look like a complete range, and
the range is annotated when either side carries a BENCHMARK_DIRECT_NO_CUDAGRAPH
marker, since a fallback row was timed under a different regime.
ab_rlc_test.sh is the same caliber as ab_fold_test.sh with the comparison axis
moved to FLAGTREE_<VENDOR>_RLC_ENHANCE. Both are caliber examples, not
regression scripts: if the switch under test no longer exists the script still
runs every round and prints a full table showing no difference, so confirm the
axis is live before reusing either.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two failure modes that cost this harness real results, both of them silent.
A wrong path does not fail as a wrong path. pytest cannot collect a benchmark
directory that is not there, so benchmark/conftest.py never registers --level
and --mode, and the run dies with "unrecognized arguments: --level core --mode
kernel" -- an error about the flags, several layers away from the missing
directory. run_pytest.sh now checks the FlagGems directory, the benchmark file,
SHAPE_FILE and REPLAY_FROM before starting pytest, prints
"PREFLIGHT_FAILED check=<the named check>" and exits 3, distinct from pytest's
own 1 and 4 so the batch layer can tell a broken environment from a broken
operator. Only the file half of a node is checked; a missing ::test_<fn> already
produces a clear "not found: ..." from pytest and needs no help.
FlagGems keeps libtuner winners in a sqlite DB under $FLAGGEMS_CACHE_DIR
(default ~/.flaggems), shared by every run on the machine. Whichever side of an
A/B runs first therefore hands its winners to the second, which then compiles
only those: measured on router_gemm, the second of two shared-cache runs
compiled 10 variants instead of 234 and finished in 17s instead of 52s. That is
how an mm A/B in this repo produced "945 variants -> 33, 136 converts -> 10" and
read as a compiler result when it was cache inheritance; isolated, both runs
compile 234.
ZL_ISOLATE_TUNER_CACHE=1 gives a run its own cache directory, dropped afterwards
with the Triton one since the chosen configs are already in autotune_records/.
It defaults to 0, because cold tuning is expensive and the standard REPLAY_FROM
A/B pins configs anyway; only record-against-record is exposed. The default is
unchanged, but the caliber is no longer invisible: either way it goes into the
run header and run_identity.json as tuner_cache, which is what was missing when
the mm numbers could not be explained after the fact.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
supports_cuda_graph was False for MUSA, so every run there was pinned to
timer=direct. That was correct when it was written -- the capture path named
torch.cuda and could not have worked -- but it is not correct now, and the two
timers are not interchangeable: direct timing walks the full Python launch path
per iteration, so on small kernels it measures mostly launch overhead. Measured
on an S5000, capture against direct: batch_norm 16x16x8x48 0.0758ms -> 0.0089ms,
tanh and gelu at 4096x4096 about 3.3x, GEMMs unchanged at 1.00x. Leaving it off
discards that on every operator that can capture, and it also puts MUSA on a
different caliber from the CUDA numbers the harness is meant to compare against.
torch_musa supplies Stream, Event, synchronize, graph and MUSAGraph, so capture
works through the device interface. Verified on an S5000 over five operators,
one per path:
tanh captured, no marker
dropout 15/15 rows fall back: MUSAGeneratorImpl::current_seed is not
callable during capture
rand same
perf_sort 25/25 rows fall back: operation not permitted when stream is
capturing
unique2 denied up front; rc=0 with 10 rows
MUSA gets its own error wording; none of the CUDA markers match it. "MUSA
error: unknown error" during capture is classified as poisoning rather than
recoverable, because the fallback runs and the process then dies in teardown:
before the deny entry was keyed correctly, unique2 reproduced exactly that,
ending at rc=1 with no rows, but now under an explicit BENCHMARK_CAPTURE_POISONED
marker instead of a bare ExceptionGroup.
capture_deny_ops is keyed by FLAGGEMS_PERF_CURRENT_OP, which holds the pytest
FUNCTION name minus `test_` -- unique2, perf_sort -- not the benchmark file
name. Getting that wrong fails open silently, which is how the first attempt
still hit rc=1 with the entry written as "unique"; a test now pins the key
space. The plugin-plan test is renamed because its assertion that
_cudagraph_plugin is absent from the plan holds on CUDA too and says nothing
about capture support.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two things kept graph-capture timing tied to NVIDIA, and both are mechanism
rather than policy, so neither backend's behaviour changes here.
The capture path named torch.cuda directly for stream, event, synchronize and
the graph type. FlagTree's do_bench already resolves those through
driver.active.get_device_interface(); its do_bench_cudagraph does not. On a
non-CUDA device every capture therefore raised and every row silently took the
fallback while the log still reported capture as enabled. The path now takes
those calls from the interface, with _graph_class picking CUDAGraph or
MUSAGraph. CUDA still calls upstream's do_bench_cudagraph verbatim, so its
numbers do not move; the port only matters where upstream cannot run.
Failure classification moves onto DeviceBackend, because the text it matches is
vendor wording that has no meaning on another runtime, and gains a third
verdict. A kernel refusing capture is ordinary and falls back. A kernel failing
on the device is fatal and re-raises, since a fallback latency there would
report a broken kernel as a measurement. New: the capture ATTEMPT can leave the
context unusable, and falling back then measures inside a dead process -- the
observed shape of that is a fallback marker followed by a teardown crash and
zero rows. Those raise under BENCHMARK_CAPTURE_POISONED instead, naming the
operator as capture-incompatible rather than broken. capture_deny_ops skips the
attempt entirely for operators already known to do this, because for them the
attempt is the damage and rediscovering it per shape costs the run.
CUDA's marker list is carried over unchanged and no poisoning wording is
guessed for it. The scratch-RNG guard around capture is skipped when an
interface exposes no default_generators, since there is then no registration to
protect.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The sharded autotune sweep could publish a config set with holes, and a run
using it reports itself as replaying while every uncovered key silently
self-tunes. That is the exact failure the record/replay caliber exists to
prevent, and it was reachable two ways.
Shape lookup was op_name-only, but BLAS operators keep their shapes under the
class name: core_shapes.yaml has BlasBenchmark, not mm. Every BLAS operator
therefore resolved to zero shapes and the plugin disabled itself with a message
that read like a normal skip. It now walks the benchmark class chain the way
FlagGems' set_shapes does. The chain is recovered statically -- the class comes
from the op_name= keyword of the constructor call in the test module, its bases
from that module plus benchmark/base.py -- because the plugin runs at
pytest_configure, before any Benchmark instance exists. Resolved keys are named
in the log, and a total miss lists the candidates it tried.
Two cases can still produce partial coverage, and both now refuse rather than
shard a subset:
- --level comprehensive, where set_shapes() merges set_more_shapes() on top of
the yaml. That addition is arbitrary per-subclass Python and invisible here.
Measured on mm: 5 of 29 shapes swept, 90 AUTOTUNE_REPLAY_FALLBACK
reason=key_missing in the run that consumed the result.
- a shard exiting non-zero, whose shapes were never swept. This used to append a
WARNING to a success message and publish the merge anyway; observed with 3 of
5 shards dead. The whole sweep is now discarded, worker logs kept.
Both fall back to the normal serial sweep, which costs wall time rather than the
same-config premise.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
run_pytest.sh hardcoded --level core, so the only shapes ever reachable were
each operator's core_shapes entry. FlagGems merges a subclass's
set_more_shapes() on top of that at --level comprehensive, and for BLAS that
addition is where the large GEMMs live: mm goes from 5 shapes to 29, taking in
four large-K tuples plus models_shapes/*.yaml, with N/K up to 152064. The
comprehensive path also yields each shape a second time (transposed, for BLAS),
so the case count doubles. None of that was reachable without editing the driver.
BENCH_LEVEL defaults to core, so existing runs are unchanged, and rejects
anything other than core/comprehensive rather than passing it through to pytest.
The chosen level is echoed in the run header, written into the archived
shapes.yaml, and recorded in run_identity.json, because an A/B whose two sides
used different levels is not comparable and is otherwise invisible afterwards.
Sweep cost scales with the shape count, so comprehensive on a BLAS operator
wants PARALLEL_WARMUP_GPUS.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Everything about which accelerator a run targets was previously spelled CUDA:
device discovery went through nvidia-smi, pinning through CUDA_VISIBLE_DEVICES,
the vendor guard only recognised NVIDIA, and the -p plugin list was hardcoded in
run_pytest.sh with no way to drop the CUDA-only members. Running the same
operator corpus on another vendor's device therefore meant editing the drivers.
The selection now hangs off one ZL_BACKEND variable (cuda/musa/rocm/ascend, with
the usual aliases):
- backends/ owns per-backend device enumeration, visibility env keys and
capability flags, so nothing above it names a vendor;
- plugin_plan.py derives the -p list from those capabilities instead of a
hardcoded array, which is what lets a non-CUDA run omit the graph-capture and
cluster-fix plugins rather than watch them fail;
- _timing_plugin selects the timer from the same capabilities: CUDA keeps
graph capture with its per-shape fallback, other backends stay on upstream
do_bench and record timer=direct, because capture-then-fallback on a device
without graph support only produces noise;
- _device_setup_plugin replaces the NVIDIA-only vendor guard;
- providers/flaggems.py resolves the pytest node, so a caller may pass TEST_NODE
directly instead of the OP/OP_FILE pair, and SHAPE_MODE=upstream can defer to
FlagGems' own core_shapes;
- _run_identity.py writes run_identity.json next to each run, recording backend,
device, timer, autotune mode and the resolved test node. Two runs whose
calibers differ are otherwise indistinguishable after the fact.
run_batch gains --backend and --devices (--gpus kept as an alias) and carries
backend/device/timer into ops_status.csv. test_backends.py covers the registry,
alias resolution, visibility env and plugin planning.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Everything in this repository runs on Linux accelerator hosts, and campaign
receipts record the SHA-256 of the harness files that produced them. With Git's
default autocrlf on Windows a checkout rewrites the shell drivers with CRLF,
which breaks them outright on the target host and changes every recorded digest.
The two inventory CSVs are renormalized here because their content is unchanged;
the remaining files normalize as part of the commits that actually modify them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Aharrypotter
