add note about memory consumption on tesla CI runner for failing test

a-r-r-o-w · a-r-r-o-w · commit 719e8d39bc7e · 2025-01-22T15:02:51.000+01:00
diff --git a/tests/models/test_modeling_common.py b/tests/models/test_modeling_common.py
@@ -1422,6 +1422,8 @@ def get_memory_usage(storage_dtype, compute_dtype):
         )
 
         self.assertTrue(fp8_e4m3_bf16_memory_footprint < fp8_e4m3_fp32_memory_footprint < fp32_memory_footprint)
+        # NOTE: the following assertion will fail on our CI (running Tesla T4) due to bf16 using more memory than fp32.
+        # On other devices, such as DGX (Ampere) and Audace (Ada), the test passes.
         self.assertTrue(fp8_e4m3_bf16_max_memory < fp8_e4m3_fp32_max_memory)
         # On this dummy test case with a small model, sometimes fp8_e4m3_fp32 max memory usage is higher than fp32 by a few
         # bytes. This only happens for some models, so we allow a small tolerance.

Original file line number	Diff line number	Diff line change
`@@ -1422,6 +1422,8 @@ def get_memory_usage(storage_dtype, compute_dtype):`
`1422`	`1422`	`)`
`1423`	`1423`
`1424`	`1424`	`self.assertTrue(fp8_e4m3_bf16_memory_footprint < fp8_e4m3_fp32_memory_footprint < fp32_memory_footprint)`
	`1425`	`+ # NOTE: the following assertion will fail on our CI (running Tesla T4) due to bf16 using more memory than fp32.`
	`1426`	`+ # On other devices, such as DGX (Ampere) and Audace (Ada), the test passes.`
`1425`	`1427`	`self.assertTrue(fp8_e4m3_bf16_max_memory < fp8_e4m3_fp32_max_memory)`
`1426`	`1428`	`# On this dummy test case with a small model, sometimes fp8_e4m3_fp32 max memory usage is higher than fp32 by a few`
`1427`	`1429`	`# bytes. This only happens for some models, so we allow a small tolerance.`