Skip to content

sycl: fuse the PQ2_0/PTQ1_0 XMX mat-muls with their neighbours (depends on #327) #328

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
kiljoy001 wants to merge 3 commits into
base: prism
Choose a base branch
from
+1,666 −295
Open

sycl: fuse the PQ2_0/PTQ1_0 XMX mat-muls with their neighbours (depends on #327) #328

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
fddc258
sycl: PQ2_0/PTQ1_0 XMX kernels from TernSYCL, PTQ1_0 decoded in-kernel
kiljoy001 Oct 7, 2026
f0a2200
sycl: fuse the PQ2_0/PTQ1_0 XMX mat-muls with their neighbours
kiljoy001 Oct 7, 2026
b61a071
sycl: PTQ1_0 XMX: tiles for small batches, decode in registers when n…
kiljoy001 Oct 7, 2026
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
14 changes: 13 additions & 1 deletion ggml/src/ggml-sycl/common.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -224,7 +224,7 @@ inline dpct::err0 ggml_sycl_set_device(const int device) try {
//////////////////////
struct optimize_feature {
bool reorder=false;
bool xmx_pq2=false; // PQ2_0 rewritten into the XMX layout (pq2_xmx.hpp); only that path can read it
bool xmx_pq2=false; // PQ2_0/PTQ1_0 rewritten into the XMX layout (pq2_xmx.hpp); only that path can read it
};

struct sycl_device_info {
Expand Down Expand Up @@ -440,12 +440,17 @@ struct ggml_backend_sycl_context {

std::unique_ptr<ggml_sycl_pool> host_pools[GGML_SYCL_MAX_DEVICES];

// XMX activations shared across mat-muls are freed out of order, which the VMM pool does not allow
std::unique_ptr<ggml_sycl_pool> xmx_act_pools[GGML_SYCL_MAX_DEVICES];

std::vector<mmid_row_mapping> mmid_row_mapping_host;

static std::unique_ptr<ggml_sycl_pool> new_pool_for_device(queue_ptr qptr, int device);

static std::unique_ptr<ggml_sycl_pool> new_pool_for_host(queue_ptr qptr, int device);

static std::unique_ptr<ggml_sycl_pool> new_unordered_pool_for_device(queue_ptr qptr, int device);

static std::unique_ptr<ggml_sycl_fattn_kv_buffers> new_fattn_kv_buffers(queue_ptr qptr, int device);

ggml_sycl_pool & pool(int device) {
Expand All @@ -459,6 +464,13 @@ struct ggml_backend_sycl_context {
return pool(device);
}

ggml_sycl_pool & xmx_act_pool() {
if (xmx_act_pools[device] == nullptr) {
xmx_act_pools[device] = new_unordered_pool_for_device(stream(device, 0), device);
}
return *xmx_act_pools[device];
}

ggml_sycl_fattn_kv_buffers & fattn_buffers(int device) {
if (fattn_bufs[device] == nullptr) {
fattn_bufs[device] = new_fattn_kv_buffers(stream(device, 0), device);
Expand Down
Loading