Support inverse transform ue8m0 scale (#13285)

test/srt/run_suite.py

@@ -75,6 +75,7 @@ suites = {
         TestFile("test_eval_fp8_accuracy.py", 303),
         TestFile("test_fa3.py", 420),
         TestFile("test_flashmla.py", 230),
+        TestFile("test_fp8_utils.py", 5),
         TestFile("rotary_embedding/test_mrope.py", 10),
         TestFile("test_function_call_parser.py", 10),
         TestFile("test_fused_moe.py", 80),

test/srt/test_fp8_utils.py

@@ -0,0 +1,44 @@
+import unittest
+
+import torch
+
+from sglang.srt.layers.quantization.fp8_utils import (
+    inverse_transform_scale_ue8m0,
+    quant_weight_ue8m0,
+    transform_scale_ue8m0,
+)
+from sglang.test.test_utils import CustomTestCase
+
+
+class TestInverseTransformScaleUe8m0(CustomTestCase):
+    def test_round_trip(self):
+        for _ in range(100):
+            weight_bf16 = torch.randn(
+                # DeepSeek V3 kv_b_proj
+                (32768, 512),
+                dtype=torch.bfloat16,
+                device="cuda",
+            )
+
+            weight_block_size = [128, 128]
+
+            qweight, sf_fp32_original = quant_weight_ue8m0(
+                weight_bf16, weight_block_size=weight_block_size
+            )
+            mn = qweight.shape[-2]
+
+            sf_packed_original = transform_scale_ue8m0(sf_fp32_original, mn=mn)
+            sf_fp32_recreated = inverse_transform_scale_ue8m0(sf_packed_original, mn=mn)
+
+            sf_packed_recreated = transform_scale_ue8m0(sf_fp32_recreated, mn=mn)
+
+            assert torch.all(
+                sf_packed_original == sf_packed_recreated
+            ), f"{sf_packed_original=} {sf_packed_recreated}"
+            assert torch.all(
+                sf_fp32_original == sf_fp32_recreated
+            ), f"{sf_fp32_original=} {sf_fp32_recreated}"
+
+
+if __name__ == "__main__":
+    unittest.main()