Repository navigation
fix failed test cases for glm5_next - #48497
Conversation
Signed-off-by: kaixuanliu <kaixuan.liu@intel.com>
Signed-off-by: kaixuanliu <kaixuan.liu@intel.com>
| @unittest.skip("Glm5Next text DSA attention does not support flash_attention_2 in the common full-model test.") | ||
| def test_flash_attn_2_inference_equivalence(self): | ||
| pass | ||
|
|
||
| @unittest.skip("Glm5Next text DSA attention does not support flash_attention_2 in the common full-model test.") | ||
| def test_flash_attn_2_inference_equivalence_right_padding(self): | ||
| pass | ||
|
|
There was a problem hiding this comment.
Hmm is it because vision side does support it but text doesnt? I'm a bit confused because the common test imo should detect to skip this by itself
The 0 division error is something else tho, probably our prepare head dim function (which we should ignore in this case as it's a unique case)
There was a problem hiding this comment.
Good point, the explicit skips were too broad. The real issue was that the common FA2 test calls _prepare_config_headdim before it instantiates the model and reaches the existing _supports_flash_attn skip logic. For GLM5 Next, text_config.head_dim intentionally mirrors qk_rope_head_dim, which is 0 in this test config, while the actual QK dim is qk_head_dim.
I replaced the manual skips with a model-test _prepare_config_headdim override that temporarily uses text_config.qk_head_dim, so the common helper no longer hits the 0-division and the common FA2 skip logic handles this as expected.
|
[For maintainers] Suggested jobs to run (before merge) run-slow: glm5_next |
CI recapDashboard: View test results in Grafana |
vasqu
left a comment
vasqu
left a comment
There was a problem hiding this comment.
Nice, thanks for fixing 🫡
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
* fix failed test cases for glm5_next Signed-off-by: kaixuanliu <kaixuan.liu@intel.com> * update comment Signed-off-by: kaixuanliu <kaixuan.liu@intel.com> * update Signed-off-by: kaixuanliu <kaixuan.liu@intel.com> * update Signed-off-by: kaixuanliu <kaixuan.liu@intel.com> --------- Signed-off-by: kaixuanliu <kaixuan.liu@intel.com>
Fix 3 failed cases for glm5-next:
@ydshieh pls help review, thx!