fix bug in quantized_linear.py when d2s and apply_rotary_embedding bug in mmha_util.cu.h (#58134)

zhoutianzi666 · web-flow · commit 973f6d09fed5 · 2023-10-17T13:49:49.000+08:00
diff --git a/paddle/phi/kernels/fusion/gpu/mmha_util.cu.h b/paddle/phi/kernels/fusion/gpu/mmha_util.cu.h
@@ -1325,16 +1325,18 @@ inline __device__ void apply_rotary_embedding(uint2& q,      // NOLINT
   k.y = rotary_embedding_transform(k.y, cos.y, sin.x);
 }
 
-inline __device__ void apply_rotary_embedding(uint2& q,       // NOLINT
-                                              uint2& k,       // NOLINT
-                                              float4& cos,    // NOLINT
-                                              float4& sin) {  // NOLINT
+inline __device__ void apply_rotary_embedding(
+    uint2& q,       // NOLINT equals 4 half.
+    uint2& k,       // NOLINT
+    float4& cos,    // NOLINT 2 float2 cos.
+    float4& sin) {  // NOLINT
   Float4_& cos_ = *reinterpret_cast<Float4_*>(&cos);
   Float4_& sin_ = *reinterpret_cast<Float4_*>(&sin);
+  // cos_.x is float2
   q.x = rotary_embedding_transform(q.x, cos_.x, sin_.x);
   k.x = rotary_embedding_transform(k.x, cos_.x, sin_.x);
   q.y = rotary_embedding_transform(q.y, cos_.y, sin_.y);
-  k.y = rotary_embedding_transform(k.y, cos_.y, sin_.x);
+  k.y = rotary_embedding_transform(k.y, cos_.y, sin_.y);
 }
 
 inline __device__ void apply_rotary_embedding(uint4& q,      // NOLINT
diff --git a/python/paddle/nn/quant/quantized_linear.py b/python/paddle/nn/quant/quantized_linear.py
@@ -164,7 +164,7 @@ def weight_only_linear(
             'weight': [weight],
             'weight_scale': [weight_scale],
         }
-        if bias:
+        if bias is not None:
             inputs["bias"] = [bias]
         attrs = {'weight_dtype': weight_dtype}
 

Original file line number	Diff line number	Diff line change
`@@ -164,7 +164,7 @@ def weight_only_linear(`
`164`	`164`	`'weight': [weight],`
`165`	`165`	`'weight_scale': [weight_scale],`
`166`	`166`	`}`
`167`		`- if bias:`
	`167`	`+ if bias is not None:`
`168`	`168`	`inputs["bias"] = [bias]`
`169`	`169`	`attrs = {'weight_dtype': weight_dtype}`
`170`	`170`