fix

orca-zhang · orca-zhang · commit 3f960653a7c8 · 2025-02-27T17:40:40.000+08:00
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -595,6 +595,7 @@ static struct ggml_tensor * llm_build_kqv(
             padded_v = ggml_pad(ctx, v, 0, k->ne[0] - v->ne[1], 0, 0);
             cb(padded_v, "padded_v", il);
             n_embd_head_v_out = n_embd_head_k;
+            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
`@@ -595,6 +595,7 @@ static struct ggml_tensor * llm_build_kqv(`
`595`	`595`	`padded_v = ggml_pad(ctx, v, 0, k->ne[0] - v->ne[1], 0, 0);`
`596`	`596`	`cb(padded_v, "padded_v", il);`
`597`	`597`	`n_embd_head_v_out = n_embd_head_k;`
	`598`	`+ padded_v = ggml_cont(ctx, padded_v);`
`598`	`599`	`}`
`599`	`600`
`600`	`601`	`cur = ggml_flash_attn_ext(ctx, q, k, padded_v, kq_mask, kq_scale, hparams.f_max_alibi_bias,`