Align Frontier FP8 profiling with vLLM runtime

2026-07-15 18:34:13 +08:00
parent 9c8570f36b
commit 684a2de413
3 changed files with 525 additions and 0 deletions
--- a/runs/frontier-multicase-sufficiency-v0/best_effort/frontier_fp8_profile_contract.patch
+++ b/runs/frontier-multicase-sufficiency-v0/best_effort/frontier_fp8_profile_contract.patch
@@ -0,0 +1,89 @@
+diff --git a/frontier/config/quantization_manager.py b/frontier/config/quantization_manager.py
+--- a/frontier/config/quantization_manager.py
+++ b/frontier/config/quantization_manager.py
+@@ -356,9 +356,17 @@ class QuantizationManager:
+             target_precision = self.get_precision(op_name)
+             precision_match = target_precision == profiling_precision
+             quant_match = profiling_quant_signature == expected_quant_signature
+            # A mixed-precision profiling CSV records the model's output dtype at
+            # file level.  Quantized operators in that CSV were nevertheless
+            # measured with the exact quantization scheme identified by the
+            # quant signature.  Treating those samples as BF16 and scaling them
+            # again would double-apply the FP8 speedup.
+            exact_fp8_profile = quant_match and target_precision == PrecisionType.FP8
+             with self._lock:
+-                self._operation_profiling_precision[op_name] = profiling_precision
+-                if precision_match:
+                self._operation_profiling_precision[op_name] = (
+                    target_precision if exact_fp8_profile else profiling_precision
+                )
+                if precision_match or exact_fp8_profile:
+                     self._operation_data_sources[op_name] = "profiling"
+                     self._operation_approximation_factors.pop(op_name, None)
+                     self._operation_speedup_factors.pop(op_name, None)
+diff --git a/tests/unit/test_quantization_profile_contract.py b/tests/unit/test_quantization_profile_contract.py
+new file mode 100644
+--- /dev/null
+++ b/tests/unit/test_quantization_profile_contract.py
+@@ -0,0 +1,61 @@
+from unittest.mock import MagicMock
+
+import pytest
+
+from frontier.config.model_config import QuantizationConfig
+from frontier.config.precision_type import PrecisionType
+from frontier.config.quantization_manager import QuantizationManager
+
+
+def test_exact_fp8_quant_signature_does_not_rescale_mixed_profile() -> None:
+    manager = QuantizationManager()
+    manager.load_config()
+
+    quant_config = QuantizationConfig(
+        quant_method="fp8",
+        activation_scheme="dynamic",
+        is_checkpoint_fp8_serialized=True,
+        weight_block_size=(128, 128),
+    )
+    quant_signature = quant_config.get_quant_signature()
+    model_config = MagicMock()
+    model_config.get_default_precision.return_value = PrecisionType.BF16
+    model_config.get_name.return_value = "Qwen3-235B-A22B-FP8"
+    model_config.torch_dtype = "bfloat16"
+    model_config.quantization_config = quant_config
+    model_config.get_quant_signature.return_value = quant_signature
+
+    manager.configure_from_model_config(model_config)
+    manager.register_profiling_metadata(
+        operation_names=["attn_pre_proj"],
+        profiling_precision=PrecisionType.BF16,
+        profiling_quant_signature=quant_signature,
+        expected_quant_signature=quant_signature,
+        file_path="linear_op.csv",
+    )
+
+    metadata = {
+        item["operation"]: item
+        for item in manager.get_operation_precision_metadata()
+    }
+    assert metadata["attn_pre_proj"]["data_source"] == "profiling"
+    assert metadata["attn_pre_proj"]["approximation_factor"] is None
+    assert manager.has_precision_mismatch("attn_pre_proj") is False
+    assert manager.adjust_compute_time("attn_pre_proj", 1.25) == pytest.approx(1.25)
+
+
+def test_mismatched_fp8_quant_signature_still_uses_approximation() -> None:
+    manager = QuantizationManager()
+    manager.load_config()
+    manager._operation_precisions["attn_pre_proj"] = PrecisionType.FP8
+
+    manager.register_profiling_metadata(
+        operation_names=["attn_pre_proj"],
+        profiling_precision=PrecisionType.BF16,
+        profiling_quant_signature="none",
+        expected_quant_signature="method=fp8",
+        file_path="linear_op.csv",
+    )
+
+    assert manager.has_precision_mismatch("attn_pre_proj") is True
+    assert manager.adjust_compute_time("attn_pre_proj", 1.0) == pytest.approx(0.5)