Use scaled mm for untuned fp8 gemm (#50)

* update quark quantizer command * typo * Using scaled_mm for untuned gemm * remove comment * fix yapf
ROCm · Jun 14, 2024 · 0438499 · 0438499
1 parent d3da246
commit 0438499
Showing 1 changed file with 33 additions and 37 deletions.
diff --git a/vllm/model_executor/layers/quantization/fp8_rocm.py b/vllm/model_executor/layers/quantization/fp8_rocm.py
@@ -24,10 +24,8 @@
 class Fp8RocmConfig(QuantizationConfig):
 
     def __init__(self) -> None:
-        # self.quantized_weights_path = config["quantized_weights"]
         self._tuned = {}
         gemm_type = os.getenv("FP8_GEMM", "fp8_16")
-        #print(f"Integral Cross factor = {self.factor}")
         if gemm_type == "fp8_8":
             self.gemm_method = Fp8RocmLinearMethod.apply_fp8_8
             tuned_filename = "/tmp/tuned_fp8_8.csv"
@@ -220,23 +218,15 @@ def apply_fp8_16(
 
         algo = self._config._tuned.get((m, n, k))
         if algo is None:
-            import os
-
-            if os.getenv("TUNE_FP8") == "1":
-                try:
-                    df = pd.read_csv("/tmp/fp8_shapes.csv")
-                except (IOError, pd.errors.EmptyDataError,
-                        pd.errors.ParserError):
-                    df = pd.DataFrame(columns=["M", "N", "K"])
-                df = pd.concat(
-                    [df, pd.DataFrame({
-                        "M": [m],
-                        "N": [n],
-                        "K": [k]
-                    })]).drop_duplicates()
-                df.to_csv("/tmp/fp8_shapes.csv", index=False)
-            algo = 0
-        res = vllm_ops.fp8_gemm_16(x8, weight.t(), asf, wsf, int(algo))
+            _save_shape(m, n, k)
+            res, _ = torch._scaled_mm(x8,
+                                      weight.t(),
+                                      out_dtype=x.dtype,
+                                      scale_a=asf,
+                                      scale_b=wsf,
+                                      bias=bias)
+        else:
+            res = vllm_ops.fp8_gemm_16(x8, weight.t(), asf, wsf, int(algo))
         return res
 
     def apply_fp8_8(
@@ -257,24 +247,16 @@ def apply_fp8_8(
 
         algo = self._config._tuned.get((m, n, k))
         if algo is None:
-            import os
-
-            if os.getenv("TUNE_FP8") == "1":
-                try:
-                    df = pd.read_csv("/projects/fp8_shapes.csv")
-                except (IOError, pd.errors.EmptyDataError,
-                        pd.errors.ParserError):
-                    df = pd.DataFrame(columns=["M", "N", "K"])
-                df = pd.concat(
-                    [df, pd.DataFrame({
-                        "M": [m],
-                        "N": [n],
-                        "K": [k]
-                    })]).drop_duplicates()
-                df.to_csv("/tmp/fp8_shapes.csv", index=False)
-            algo = 0
-
-        res = vllm_ops.fp8_gemm(x8, weight.t(), asf, wsf, osf, int(algo))
+            _save_shape(m, n, k)
+            res, _ = torch._scaled_mm(x8,
+                                      weight.t(),
+                                      out_dtype=x8.dtype,
+                                      scale_a=asf,
+                                      scale_b=wsf,
+                                      scale_result=osf,
+                                      bias=bias)
+        else:
+            res = vllm_ops.fp8_gemm(x8, weight.t(), asf, wsf, osf, int(algo))
         res16 = torch.empty_like(res, dtype=torch.float16)
         vllm_ops.convert_fp8(res16, res, 1 / osf)
         return res16
@@ -308,3 +290,17 @@ def _per_tensor_dequantize(tensor: torch.Tensor,
     fake_qweight = tensor.to(torch.float16)
     dq_weight = fake_qweight * inv_scale
     return dq_weight
+
+
+def _save_shape(m, n, k):
+    if os.getenv("TUNE_FP8") == "1":
+        try:
+            df = pd.read_csv("/tmp/fp8_shapes.csv")
+        except (IOError, pd.errors.EmptyDataError, pd.errors.ParserError):
+            df = pd.DataFrame(columns=["M", "N", "K"])
+        df = pd.concat([df, pd.DataFrame({
+            "M": [m],
+            "N": [n],
+            "K": [k]
+        })]).drop_duplicates()
+        df.to_csv("/tmp/fp8_shapes.csv", index=False)