Skip to content

Fix CUDA KV cache clear on non-host buffers #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
double-thinker wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix CUDA KV cache clear on non-host buffers #4

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
34 changes: 34 additions & 0 deletions src/voxtral.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1076,6 +1076,31 @@ static void clear_kv_cache(voxtral_context * ctx) {
if (!ctx || !ctx->kv_self_k || !ctx->kv_self_v) {
return;
}
const bool k_host = ctx->kv_self_k->buffer != nullptr &&
ggml_backend_buffer_is_host(ctx->kv_self_k->buffer);
const bool v_host = ctx->kv_self_v->buffer != nullptr &&
ggml_backend_buffer_is_host(ctx->kv_self_v->buffer);

// On non-host buffers (CUDA/Vulkan/Metal), raw memset on ggml_get_data()
// may access device pointers and segfault. Clear through backend ops instead.
if (!k_host || !v_host) {
if (ctx->buf_persistent) {
ggml_backend_buffer_clear(ctx->buf_persistent, 0);
} else {
static thread_local std::vector<uint8_t> zeros(1 << 20, 0);

for (ggml_tensor * t : {ctx->kv_self_k, ctx->kv_self_v}) {
const size_t nbytes = ggml_nbytes(t);
for (size_t off = 0; off < nbytes; off += zeros.size()) {
const size_t n = std::min(zeros.size(), nbytes - off);
ggml_backend_tensor_set(t, zeros.data(), off, n);
}
}
}
ctx->kv_used = 0;
return;
}

void * k_data = ggml_get_data(ctx->kv_self_k);
void * v_data = ggml_get_data(ctx->kv_self_v);
if (k_data) {
Expand All @@ -1091,6 +1116,15 @@ static void kv_cache_shift_left(voxtral_context * ctx, int32_t shift) {
if (!ctx || shift <= 0 || !ctx->kv_self_k || !ctx->kv_self_v) {
return;
}
const bool k_host = ctx->kv_self_k->buffer != nullptr &&
ggml_backend_buffer_is_host(ctx->kv_self_k->buffer);
const bool v_host = ctx->kv_self_v->buffer != nullptr &&
ggml_backend_buffer_is_host(ctx->kv_self_v->buffer);
if (!k_host || !v_host) {
clear_kv_cache(ctx);
return;
}

const int32_t window = VOXTRAL_DEC_WINDOW;
if (shift >= window) {
clear_kv_cache(ctx);
Expand Down