metal : batch rows copy in a single threadgroup (ggml-org#14384)

ggerganov · Minh141120 · commit aea1ecaf0034 · 2025-07-05T23:55:35.000+07:00
* metal : batch rows copy in a single threadgroup

ggml-ci

* metal : handle some edge cases when threadgroup size is not a power of 2

ggml-ci
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
@@ -2571,6 +2571,7 @@ static bool ggml_metal_encode_node(
                     nth *= 2;
                 }
 
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
                 nth = MIN(nth, ne00);
 
                 ggml_metal_kargs_sum_rows args = {

Original file line number	Diff line number	Diff line change
`@@ -2571,6 +2571,7 @@ static bool ggml_metal_encode_node(`
`2571`	`2571`	`nth *= 2;`
`2572`	`2572`	`}`
`2573`	`2573`
	`2574`	`+ nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);`
`2574`	`2575`	`nth = MIN(nth, ne00);`
`2575`	`2576`
`2576`	`2577`	`ggml_metal_kargs_sum_rows args = {`