Merge pull request #3277 from AI-Hypercomputer:agagik-qk-logits

PiperOrigin-RevId: 877549235

Author: Google-ML-Automation

src/maxtext/utils/qk_clip_utils.py

@@ -53,7 +53,8 @@ def extract_logits(path, val):
   if not all_max_logits:
     return None
 
-  return jnp.max(jnp.stack(all_max_logits))
+  # Compute max per layer first to handle potential shape mismatches
+  return jnp.max(jnp.stack([jnp.max(x) for x in all_max_logits]))
 
 
 def apply_qk_clip(state, intermediate_outputs, config):

tests/unit/train_compile_test.py

@@ -839,3 +839,31 @@ def test_engram_integration(self):
             "hf_access_token=fake",
         )
     )
+
+  @pytest.mark.cpu_only
+  def test_qk_clip(self):
+    """AOT test for qk-clip with DeepSeek3 Tiny model"""
+    compiled_trainstep_file = "/tmp/test_qk_clip.pickle"
+    train_compile_main(
+        (
+            "",
+            get_test_config_path(),
+            f"compiled_trainstep_file={compiled_trainstep_file}",
+            "compile_topology=v5p-8",
+            "compile_topology_num_slices=1",
+            "model_name=deepseek3-tiny",
+            "scan_layers=True",
+            "sparse_matmul=True",
+            "megablox=True",
+            "use_tokamax_gmm=False",
+            # TODO(agagik): update to flash after support
+            "attention=dot_product",
+            "use_tokamax_splash=True",
+            "max_target_length=128",
+            "per_device_batch_size=1",
+            "dtype=bfloat16",
+            "weight_dtype=float32",
+            "use_qk_clip=true",
+            "qk_clip_threshold=100",
+        )
+    )