Skip to content

cuda: DGX Spark / GB10 backend support — HBM-resident model #13

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
TrevorS wants to merge 1 commit into
base: main
Choose a base branch
from
+234 −92
Open

cuda: DGX Spark / GB10 backend support — HBM-resident model #13

Show file tree
Hide file tree
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
59 changes: 42 additions & 17 deletions ds4.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1394,6 +1394,13 @@ static uint64_t accelerator_cuda_preload_span_bytes(void) {
}

static bool accelerator_cache_model_tensor_spans(const ds4_model *m, uint64_t *cached_out) {
/* Routed MoE expert weights (`*_exps.weight`) are ~65 GiB of the model on
* V4-Flash but only top-K of N=256 experts fire per token — pre-caching
* them in HBM wastes most of the budget on cold weights and starves the
* hot non-MoE tensors that every token reads. Skip them at the span-
* build stage so the cap fills with attn / shared FFN / embedding /
* output head. Cold MoE expert reads fall back to the UVA-mapped
* pointer. */
accelerator_tensor_span *spans = xmalloc((size_t)m->n_tensors * sizeof(spans[0]));
uint64_t nspan = 0;
for (uint64_t i = 0; i < m->n_tensors; i++) {
Expand All @@ -1403,6 +1410,11 @@ static bool accelerator_cache_model_tensor_spans(const ds4_model *m, uint64_t *c
free(spans);
return false;
}
/* Routed-expert weights are the only tensors with "_exps." in the
* name; memmem is safe on short names (returns NULL). */
if (memmem(t->name.ptr, t->name.len, "_exps.", 6) != NULL) {
continue;
}
spans[nspan++] = (accelerator_tensor_span){
.off = t->abs_offset,
.end = t->abs_offset + t->bytes,
Expand Down Expand Up @@ -13183,11 +13195,9 @@ static bool metal_graph_eval_token_raw_swa_top(
if (ok) ok = metal_graph_encode_token_raw_swa(g, model, weights,
token, pos, true, true);
if (ok) {
ok = ds4_gpu_indexer_topk_tensor(g->comp_selected,
g->logits,
DS4_N_VOCAB,
1,
1) != 0;
ok = ds4_gpu_argmax_tensor(g->comp_selected,
g->logits,
DS4_N_VOCAB) != 0;
}
if (ok) ok = ds4_gpu_end_commands() != 0;
if (ok) ok = ds4_gpu_tensor_read(g->comp_selected, 0, top_id, sizeof(*top_id)) != 0;
Expand Down Expand Up @@ -13294,11 +13304,9 @@ static bool metal_graph_eval_mtp_draft_from_hc(
mtp,
base_weights->output->dim[1]);
if (ok && top_id) {
ok = ds4_gpu_indexer_topk_tensor(g->comp_selected,
g->logits,
DS4_N_VOCAB,
1,
1) != 0;
ok = ds4_gpu_argmax_tensor(g->comp_selected,
g->logits,
DS4_N_VOCAB) != 0;
}
if (ok) ok = ds4_gpu_end_commands() != 0;
g->cur_hc = saved_cur;
Expand Down Expand Up @@ -14085,7 +14093,14 @@ static bool metal_graph_verify_suffix_tops(
n_tokens,
weights->output->dim[1]);
if (ok) {
if (top_rows) {
if (top_rows == 1) {
/* Common K=2 verify case: top_k=1 over n_vocab → use the dedicated
* argmax kernel (single-block tree-reduce) instead of the legacy
* indexer_topk_kernel's single-thread O(n_vocab * top_k) fall-through. */
ok = ds4_gpu_argmax_tensor(g->comp_selected,
g->spec_logits,
DS4_N_VOCAB) != 0;
} else if (top_rows) {
ok = ds4_gpu_indexer_topk_tensor(g->comp_selected,
g->spec_logits,
DS4_N_VOCAB,
Expand Down Expand Up @@ -14216,11 +14231,9 @@ static bool metal_graph_verify_decode2_exact(
g->cur_hc = cur0;
ok = ds4_gpu_begin_commands() != 0;
if (ok) ok = metal_graph_encode_output_head(g, model, weights, weights->output->dim[1]);
if (ok) ok = ds4_gpu_indexer_topk_tensor(g->comp_selected,
g->logits,
DS4_N_VOCAB,
1,
1) != 0;
if (ok) ok = ds4_gpu_argmax_tensor(g->comp_selected,
g->logits,
DS4_N_VOCAB) != 0;
if (ok) ok = ds4_gpu_end_commands() != 0;
else (void)ds4_gpu_synchronize();
g->cur_hc = saved_cur;
Expand Down Expand Up @@ -17735,13 +17748,25 @@ int ds4_engine_open(ds4_engine **out, const ds4_engine_options *opt) {
*out = NULL;
return 1;
}
if (!e->mtp_ready && !accelerator_cache_model_tensors(e->backend, &e->model)) {
if (!accelerator_cache_model_tensors(e->backend, &e->model)) {
fprintf(stderr, "ds4: %s failed to prepare startup model cache\n",
ds4_backend_name(e->backend));
ds4_engine_close(e);
*out = NULL;
return 1;
}
/* Also populate the HBM cache for the MTP support model when loaded.
* Without this, MTP-block tensor reads at decode time hit the UVA-
* mapped pointer (slow) instead of cudaMalloc'd HBM copies (fast).
* The MoE expert filter in accelerator_cache_model_tensor_spans
* skips `mtp.0.ffn_*_exps.weight` automatically. */
if (e->mtp_ready && !accelerator_cache_model_tensors(e->backend, &e->mtp_model)) {
fprintf(stderr, "ds4: %s failed to prepare MTP startup model cache\n",
ds4_backend_name(e->backend));
ds4_engine_close(e);
*out = NULL;
return 1;
}
fprintf(stderr, "ds4: %s backend initialized for graph diagnostics\n",
ds4_backend_name(e->backend));
}
Expand Down
Loading