llama : move no_alloc out of llama_hparams (#31141)

* llama : move no_alloc out of llama_hparams

Assisted-by: dsh:llama.cpp/ggml-org/GLM-5.3-Flash-GGUF (2x DGX Spark (RDMA, TP=2))

* cont : clean-up

* cont : clean-up

Assisted-by: dsh:llama.cpp/ggml-org/GLM-5.3-Flash-GGUF (2x DGX Spark (RDMA, TP=2))

* tests : add a no_alloc case to test-llama-context

Assisted-by: dsh:llama.cpp/ggml-org/GLM-5.3-Flash-GGUF (2x DGX Spark (RDMA, TP=2))
This commit is contained in:
Georgi Gerganov authored and GitHub committed 2026-10-26 16:21:25 +02:00
1 parent 4b851fbee8
commit a4699400e3
30 files changed
+266 -179

No files matched your search

+52
View File
@@ -4,6 +4,8 @@
#include "llama-io.h"
#include <set>
#include <stdexcept>
#include <string>
#include <vector>
llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1) {
@@ -64,6 +66,52 @@ bool llama_memory_status_is_fail(llama_memory_status status) {
return false;
}
//
// llama_memory_bufs
//
ggml_backend_buffer_ptr llama_memory_bufs::alloc_buf(ggml_context * ctx, ggml_backend_buffer_type_t buft, bool no_alloc, const char * name) {
ggml_backend_buffer_ptr buf;
if (no_alloc) {
buf.reset(ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0)); // dummy buffer
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
t->buffer = buf.get(); // set dummy buffer so that the backend scheduler won't try to allocate the tensors
}
} else {
buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft)); // real buffer
}
if (!buf) {
throw std::runtime_error(std::string("failed to allocate buffer for ") + name);
}
return buf;
}
ggml_backend_buffer_t llama_memory_bufs::emplace(ggml_context_ptr ctx, ggml_backend_buffer_type_t buft, const char * name) {
ggml_backend_buffer_ptr buf = alloc_buf(ctx.get(), buft, no_alloc, name);
// initialize the buffer to avoid NaNs in the padding
ggml_backend_buffer_clear(buf.get(), 0);
bufs.emplace_back(std::move(ctx), std::move(buf));
return bufs.back().second.get();
}
std::map<ggml_backend_buffer_type_t, size_t> llama_memory_bufs::memory_breakdown() const {
std::map<ggml_backend_buffer_type_t, size_t> ret;
for (const auto & [ctx, buf] : bufs) {
ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf.get());
if (no_alloc) {
GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) == nullptr);
ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
} else {
ret[buft] += ggml_backend_buffer_get_size(buf.get());
}
}
return ret;
}
// collects the buffers of the tensors that a state write visits, without keeping the data
struct llama_memory_buffers_writer : public llama_io_write_i {
std::set<ggml_backend_buffer_t> buffers;
@@ -84,6 +132,10 @@ struct llama_memory_buffers_writer : public llama_io_write_i {
};
std::vector<ggml_backend_buffer_t> llama_memory_get_buffers(const llama_memory_t mem) {
if (mem == nullptr) {
return {};
}
llama_memory_buffers_writer writer;
mem->state_write(writer);