graph : utilize ggml_build_forward_select() to avoid reallocations (#18898)

* graph : avoid branches between embedding and token inputs

* models : make deepstack graphs (e.g. Qwen3 VL) have constant topology

* ci : enable -DGGML_SCHED_NO_REALLOC=ON for server CI

* cont : pad token embeddings to n_embd_inp
This commit is contained in:
Georgi Gerganov authored and GitHub committed 2026-01-23 18:22:34 +02:00
1 parent cb6caca191
commit 557515be1e
7 files changed
+69 -53

No files matched your search

+1 -1
View File
@@ -2903,7 +2903,7 @@ void llama_context::opt_epoch_iter(
};
ctx_compute_opt = ggml_init(params);
}
ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_tokens(), res->get_logits());
ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_inp_tokens(), res->get_logits());
ggml_opt_alloc(opt_ctx, train);
res->set_inputs(&ubatch);