hugo/DS4Server
1
0
Fork 0
Code Issues 2 Pull Requests Actions Packages Projects Releases Wiki Activity

Establish the Mojo 1.0 integration gate

Browse Source
This commit is contained in:
Georg Bauer
2026-08-31 15:25:04 +02:00
parent 2c2c3702b0
commit 0dcf181dbe
11 changed files with 287 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
native/metal/ds4_metal.m
View File

@@ -13111,12 +13111,7 @@ static int ds4_gpu_stream_full_expert_addr_table_requested(void) {
static uint64_t ds4_gpu_buffer_address(id<MTLBuffer> buffer, NSUInteger inner) {
if (!buffer) return 0;
#if TARGET_OS_OSX
if (@available(macOS 13.0, *)) {
return (uint64_t)[buffer gpuAddress] + (uint64_t)inner;
}
#endif
return 0;
return (uint64_t)[buffer gpuAddress] + (uint64_t)inner;
}
static int ds4_gpu_stream_compact_addr_ensure_buffers(uint32_t layer) {
@@ -17286,50 +17281,42 @@ static DS4MetalQ4ExpertTable *ds4_gpu_q4_expert_address_table(
n_total_expert];
if (!expert_buffers) return nil;
#if TARGET_OS_OSX
if (@available(macOS 13.0, *)) {
for (uint32_t first = 0; first < n_total_expert; first += table_group_size) {
const uint32_t remaining = n_total_expert - first;
const uint32_t group_n =
remaining < table_group_size ? remaining : table_group_size;
if ((uint64_t)first > UINT64_MAX / expert_bytes ||
(uint64_t)group_n > UINT64_MAX / expert_bytes) {
fprintf(stderr, "ds4: Metal Q4 expert address table group byte overflow\n");
return nil;
}
const uint64_t rel = (uint64_t)first * expert_bytes;
const uint64_t group_bytes = (uint64_t)group_n * expert_bytes;
if (rel > UINT64_MAX - tensor_offset) {
fprintf(stderr, "ds4: Metal Q4 expert address table group offset overflow\n");
return nil;
}
uint64_t inner = 0;
id<MTLBuffer> group_buf = nil;
if (table_group_size > 1) {
group_buf = ds4_gpu_wrap_model_range(model_map,
model_size,
tensor_offset + rel,
group_bytes,
&inner);
} else {
group_buf = ds4_gpu_wrap_model_exact_range_owned(model_map,
model_size,
tensor_offset + rel,
expert_bytes,
&inner);
}
if (!group_buf) return nil;
const uint64_t base_address = (uint64_t)[group_buf gpuAddress] + inner;
for (uint32_t j = 0; j < group_n; j++) {
addresses[first + j] = base_address + (uint64_t)j * expert_bytes;
}
[expert_buffers addObject:group_buf];
for (uint32_t first = 0; first < n_total_expert; first += table_group_size) {
const uint32_t remaining = n_total_expert - first;
const uint32_t group_n =
remaining < table_group_size ? remaining : table_group_size;
if ((uint64_t)first > UINT64_MAX / expert_bytes ||
(uint64_t)group_n > UINT64_MAX / expert_bytes) {
fprintf(stderr, "ds4: Metal Q4 expert address table group byte overflow\n");
return nil;
}
} else
#endif
{
fprintf(stderr, "ds4: Metal GPU addresses require macOS 13 or newer\n");
return nil;
const uint64_t rel = (uint64_t)first * expert_bytes;
const uint64_t group_bytes = (uint64_t)group_n * expert_bytes;
if (rel > UINT64_MAX - tensor_offset) {
fprintf(stderr, "ds4: Metal Q4 expert address table group offset overflow\n");
return nil;
}
uint64_t inner = 0;
id<MTLBuffer> group_buf = nil;
if (table_group_size > 1) {
group_buf = ds4_gpu_wrap_model_range(model_map,
model_size,
tensor_offset + rel,
group_bytes,
&inner);
} else {
group_buf = ds4_gpu_wrap_model_exact_range_owned(model_map,
model_size,
tensor_offset + rel,
expert_bytes,
&inner);
}
if (!group_buf) return nil;
const uint64_t base_address = (uint64_t)[group_buf gpuAddress] + inner;
for (uint32_t j = 0; j < group_n; j++) {
addresses[first + j] = base_address + (uint64_t)j * expert_bytes;
}
[expert_buffers addObject:group_buf];
}
[address_buffer didModifyRange:NSMakeRange(0,