Skip to content

fix(pytorch_xdit): replace placeholder thresholds with measured values; fix cache paths - #324

Open
speriaswamy-amd wants to merge 2 commits into
speriaswamy/xdit-10-token-redactionfrom
speriaswamy/xdit-11-measured-thresholds
Open

fix(pytorch_xdit): replace placeholder thresholds with measured values; fix cache paths#324
speriaswamy-amd wants to merge 2 commits into
speriaswamy/xdit-10-token-redactionfrom
speriaswamy/xdit-11-measured-thresholds

Conversation

@speriaswamy-amd

@speriaswamy-amd speriaswamy-amd commented Aug 14, 2026

Copy link
Copy Markdown
Contributor

Part 11 of 12 in a stack that replaces #184. Base: #323.

Thresholds (AIMVT-88)

The shipped mi300x thresholds were unvalidated placeholders and one was wrong by more than an order of magnitude: wan22_i2v_a14b set max_avg_total_time_s: 12.0, while a real MI300X run with compile: true averages 161.03 s/step (5 clean steps: 160.98, 161.03, 160.98, 161.06, 161.11). Every real run on that hardware failed the threshold.

benchmark gpu before after basis
wan22_i2v_a14b mi300x 12.0 200.0 161.03 s/step measured, ~25% margin
wan22_i2v_a14b mi350 120.0 86.10–90.99 s/step across two nodes
flux1_dev_t2i mi300x 8.0 3.0 1.39 s/pipe over 25 reps, ~2x margin
flux1_dev_t2i mi350 2.0 0.92 s/pipe, consistent across two runs

auto and mi355 are unchanged and remain unvalidated. Each config gains an _expected_results_comment recording which numbers are measured, on what hardware, with what margin — and which are not.

Schema

PytorchXditWan22Benchmarks and PytorchXditFlux1DevBenchmarks use extra="forbid", so the comment field would make every test in the module fail validation before running. Both switch to extra="allow", matching the other schemas in the same file that already permit comment fields.

Cache paths

hf_home was /home/{user-id}, but HF_HOME must point at the root containing hub/ — the test builds hub/models--*/snapshots/<rev> underneath it, which does not exist directly in a home directory. Corrected to {home}/.cache/huggingface, the hf CLI default. hf_token_file and output_base_dir move from /home/{user-id}/... to {home}/..., which resolves correctly where homes are not under /home.

…s; fix cache paths

Thresholds. The shipped mi300x thresholds were unvalidated placeholders and one
was wrong by more than an order of magnitude: wan22_i2v_a14b set
max_avg_total_time_s: 12.0, while a real MI300X run with compile: true averages
161.03s/step (5 clean steps: 160.98, 161.03, 160.98, 161.06, 161.11). Every
real run on that hardware failed the threshold. Replaced with values derived
from live measurements, plus mi350 entries measured on Ruby:

  wan22_i2v_a14b  mi300x  12.0 -> 200.0  (161.03s/step measured, ~25% margin)
  wan22_i2v_a14b  mi350   new    120.0   (86.10-90.99s/step across two nodes)
  flux1_dev_t2i   mi300x   8.0 ->   3.0  (1.39s/pipe over 25 reps, ~2x margin)
  flux1_dev_t2i   mi350   new      2.0   (0.92s/pipe, consistent across 2 runs)

auto and mi355 are left as-is and remain unvalidated. Each config gains an
_expected_results_comment recording which numbers are measured, on what
hardware and with what margin, and which are not — so the next reader does not
have to guess again.

Schema. PytorchXditWan22Benchmarks and PytorchXditFlux1DevBenchmarks use
extra="forbid", so the comment field would make every test in the module fail
validation before running. Both switch to extra="allow", matching the other
schemas in the file that already permit comment fields.

Cache paths. hf_home was "/home/{user-id}", but HF_HOME must point at the root
containing hub/ — the test builds hub/models--*/snapshots/<rev> underneath it,
which does not exist directly in a home directory. Corrected to
"{home}/.cache/huggingface", the hf CLI default. hf_token_file and
output_base_dir move from "/home/{user-id}/..." to "{home}/...", which resolves
correctly where homes are not under /home.

Co-Authored-By: Claude <noreply@anthropic.com>
The _expected_results_comment fields named an internal AMD cluster;
this repo is public, so keep the measurement basis without the
cluster name.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant