Fix regression with 8 bit quant mat mul.

trivedivivek · abhinaykukkadapu · commit bafceb9f1ab2 · 2025-11-06T14:32:11.000-08:00
Differential Revision: D85767668 Pull Request resolved: pytorch#15445
diff --git a/backends/vulkan/runtime/graph/ops/glsl/linear_qcsnw_tiled.yaml b/backends/vulkan/runtime/graph/ops/glsl/linear_qcsnw_tiled.yaml
@@ -12,7 +12,7 @@ linear_qcsnw_tiled:
     WEIGHT_STORAGE: texture2d
     SCALES_STORAGE: texture2d
     TILE_ROWS: 4
-    TILE_TXCOLS: 2
+    TILE_TXCOLS: 1
     QUANT_NBITS: 8
   generate_variant_forall:
     TILE_ROWS:
diff --git a/backends/vulkan/runtime/graph/ops/impl/QuantizedLinearQCSNW.cpp b/backends/vulkan/runtime/graph/ops/impl/QuantizedLinearQCSNW.cpp
@@ -73,7 +73,7 @@ utils::uvec3 linear_qcsnw_tiled_global_wg_size(
   }
 
   // Number of output texels in the output tile
-  uint32_t out_tile_ntxcols = 2;
+  uint32_t out_tile_ntxcols = 1;
   if (quant_nbits == 4) {
     out_tile_ntxcols = 2;
   }
@@ -325,7 +325,7 @@ void add_linear_qcsnw_tiled_node(
   }
 
   // Number of output texels in the output tile
-  uint32_t out_tile_ntxcols = 2;
+  uint32_t out_tile_ntxcols = 1;
   if (quant_nbits == 4) {
     out_tile_ntxcols = 2;
   }

Original file line number	Diff line number	Diff line change
`@@ -73,7 +73,7 @@ utils::uvec3 linear_qcsnw_tiled_global_wg_size(`
`73`	`73`	`}`
`74`	`74`
`75`	`75`	`// Number of output texels in the output tile`
`76`		`- uint32_t out_tile_ntxcols = 2;`
	`76`	`+ uint32_t out_tile_ntxcols = 1;`
`77`	`77`	`if (quant_nbits == 4) {`
`78`	`78`	`out_tile_ntxcols = 2;`
`79`	`79`	`}`
`@@ -325,7 +325,7 @@ void add_linear_qcsnw_tiled_node(`
`325`	`325`	`}`
`326`	`326`
`327`	`327`	`// Number of output texels in the output tile`
`328`		`- uint32_t out_tile_ntxcols = 2;`
	`328`	`+ uint32_t out_tile_ntxcols = 1;`
`329`	`329`	`if (quant_nbits == 4) {`
`330`	`330`	`out_tile_ntxcols = 2;`
`331`	`331`	`}`