fix

orca-zhang · orca-zhang · commit 87f1435c7a6d · 2025-02-27T18:21:29.000+08:00
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -593,7 +593,6 @@ static struct ggml_tensor * llm_build_kqv(
         if (n_embd_head_v < n_embd_head_k) {
             padded_v = ggml_pad(ctx, v, 0, k->ne[0] - v->ne[1], 0, 0);
             cb(padded_v, "padded_v", il);
-            padded_v = ggml_cont(ctx, padded_v);
         }
 
         cur = ggml_flash_attn_ext(ctx, q, k, padded_v, kq_mask, kq_scale, hparams.f_max_alibi_bias,

Original file line number	Diff line number	Diff line change
`@@ -593,7 +593,6 @@ static struct ggml_tensor * llm_build_kqv(`
`593`	`593`	`if (n_embd_head_v < n_embd_head_k) {`
`594`	`594`	`padded_v = ggml_pad(ctx, v, 0, k->ne[0] - v->ne[1], 0, 0);`
`595`	`595`	`cb(padded_v, "padded_v", il);`
`596`		`- padded_v = ggml_cont(ctx, padded_v);`
`597`	`596`	`}`
`598`	`597`
`599`	`598`	`cur = ggml_flash_attn_ext(ctx, q, k, padded_v, kq_mask, kq_scale, hparams.f_max_alibi_bias,`