fix: fp8 scaling_mode is not validated before use

[andrewwhitecdw]

This PR addresses the following issue in tests/pytorch/attention/run_attention_with_cp.py: fp8 scaling_mode is not validated before use.

Changes

• tests/pytorch/attention/run_attention_with_cp.py: fp8 scaling_mode is not validated before use.

Details

--- a/tests/pytorch/attention/run_attention_with_cp.py
+++ b/tests/pytorch/attention/run_attention_with_cp.py
@@ -1,3 +1,9 @@
-    is_training = is_training == "True"
-
-    # set up environment variables and config
+    is_training = is_training == "True"
+
+    # Validate FP8 scaling mode before any conditional use so that an
+    # unsupported value fails fast with a clear message instead of
+    # leaving fp8_recipe / quantizers undefined.
+    if dtype == "fp8" and scaling_mode not in ("delayed", "current", "mxfp8"):
+        raise ValueError(f"Unsupported scaling_mode for FP8: {scaling_mode!r}")
+
+    # set up environment variables and config

Tests

• tests/pytorch/attention/test_attention_with_cp_validation.py

--- /dev/null
+++ tests/pytorch/attention/test_attention_with_cp_validation.py
@@ -0,0 +1,9 @@
+import pytest
+
+from run_attention_with_cp import run_dpa_with_cp
+
+
+def test_run_dpa_with_cp_rejects_invalid_fp8_scaling_mode():
+    """An unsupported FP8 scaling mode must raise a clear ValueError."""
+    with pytest.raises(ValueError, match="Unsupported scaling_mode for FP8"):
+        run_dpa_with_cp(dtype="fp8", scaling_mode="unsupported")

[greptile-apps]

Greptile Summary

This PR adds early validation for FP8 scaling modes in the context-parallel attention test helper.

• Accepts the three scaling modes already supported by downstream recipe and quantizer branches.
• Raises a clear ValueError for unsupported FP8 scaling modes before undefined state is used.

Confidence Score: 5/5

The PR appears safe to merge, with the new validation matching the helper’s existing supported scaling modes.

Existing callers and downstream recipe and quantizer branches consistently use only delayed, current, and mxfp8, so the added check prevents invalid state without rejecting a supported configuration.

Important Files Changed

Filename	Overview
tests/pytorch/attention/run_attention_with_cp.py	Adds fail-fast FP8 scaling-mode validation whose allowlist matches existing callers and downstream implementations.

_{Reviews (1): Last reviewed commit: "fix: fp8 scaling_mode is not validated b..." | Re-trigger Greptile}

[cyanguwa]

This path is unreachable in our test infra so the check here is probably unnecessary:
dtype == "fp8" in only tested in test_cp_with_fused_attention, and it's guarded here.

[andrewwhitecdw]

Closing this PR. @cyanguwa is right — the fp8 path is only exercised by test_cp_with_fused_attention, where scaling_mode is already parametrized to valid values and guarded (test_attention_with_cp.py:600-603), so the validation check was unreachable. Thanks for the review!