llama : reuse compute graphs #14482
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -227,8 +227,8 @@ llama_context::llama_context( | |
|
|
||
| LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes); | ||
|
|
||
| gf_res_prev.reset(new llm_graph_result(max_nodes)); | ||
| gf_res_reserve.reset(new llm_graph_result(max_nodes)); | ||
|
|
||
| // TODO: move these checks to ggml_backend_sched | ||
| // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary | ||
|
|
@@ -388,10 +388,6 @@ ggml_backend_sched_t llama_context::get_sched() const { | |
| return sched.get(); | ||
| } | ||
|
|
||
| uint32_t llama_context::n_ctx() const { | ||
| return cparams.n_ctx; | ||
| } | ||
|
|
@@ -678,38 +674,59 @@ bool llama_context::apply_adapter_cvec( | |
| return cvec.apply(model, data, len, n_embd, il_start, il_end); | ||
| } | ||
|
|
||
| llm_graph_result_i * llama_context::process_ubatch(const llama_ubatch & ubatch, llm_graph_type gtype, llama_memory_context_i * mctx, ggml_status & ret) { | ||
| if (mctx && !mctx->apply()) { | ||
| LLAMA_LOG_ERROR("%s: failed to apply memory context\n", __func__); | ||
| ret = GGML_STATUS_FAILED; | ||
| return nullptr; | ||
| } | ||
|
|
||
| auto * res = gf_res_prev.get(); | ||
| auto * gf = res->get_gf(); | ||
|
|
||
| // the new graph parameters | ||
| // in order to correctly reuse a graph, it's full topology has to be uniquely determined by these parameters | ||
| const auto gparams = graph_params(res, ubatch, mctx, gtype); | ||
|
|
||
| if (res->can_reuse(gparams)) { | ||
| //LLAMA_LOG_DEBUG("%s: reusing previous graph\n", __func__); | ||
|
|
||
| n_reused++; | ||
| } else { | ||
| res->reset(); | ||
|
|
||
| ggml_backend_sched_reset(sched.get()); | ||
| ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data); | ||
|
|
||
| //const auto t_start_us = ggml_time_us(); | ||
|
|
||
| gf = model.build_graph(gparams); | ||
|
|
||
| //LLAMA_LOG_INFO("graph build time: %.3f ms\n", (ggml_time_us() - t_start_us)/1000.0); | ||
|
|
||
| if (!gf) { | ||
| LLAMA_LOG_ERROR("%s: failed to initialize graph\n", __func__); | ||
| ret = GGML_STATUS_FAILED; | ||
| return nullptr; | ||
| } | ||
|
|
||
| if (!ggml_backend_sched_alloc_graph(sched.get(), gf)) { | ||
| LLAMA_LOG_ERROR("%s: failed to allocate graph\n", __func__); | ||
| ret = GGML_STATUS_ALLOC_FAILED; | ||
| return nullptr; | ||
| } | ||
| } | ||
|
|
||
| // set the input data for the input tensors | ||
| { | ||
| //const auto t_start_us = ggml_time_us(); | ||
|
|
||
| res->set_inputs(&ubatch); | ||
|
|
||
| //LLAMA_LOG_INFO("graph set inputs time: %.3f ms\n", (ggml_time_us() - t_start_us)/1000.0); | ||
| } | ||
|
|
||
| const auto status = graph_compute(res->get_gf(), ubatch.n_tokens > 1); | ||
| if (status != GGML_STATUS_SUCCESS) { | ||
| LLAMA_LOG_ERROR("%s: failed to compute graph, compute status: %d\n", __func__, status); | ||
| ret = status; | ||
|
|
@@ -767,9 +784,6 @@ int llama_context::encode(const llama_batch & batch_inp) { | |
|
|
||
| n_outputs = n_tokens; | ||
|
|
||
| const auto causal_attn_org = cparams.causal_attn; | ||
|
|
||
| // always use non-causal attention for encoder graphs | ||
|
|
@@ -778,7 +792,7 @@ int llama_context::encode(const llama_batch & batch_inp) { | |
| cparams.causal_attn = false; | ||
|
|
||
| ggml_status status; | ||
| const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_ENCODER, nullptr, status); | ||
|
|
||
| cparams.causal_attn = causal_attn_org; | ||
|
|
||
|
|
@@ -844,10 +858,6 @@ int llama_context::encode(const llama_batch & batch_inp) { | |
| } | ||
| } | ||
|
|
||
| // TODO: hacky solution | ||
| if (model.arch == LLM_ARCH_T5 && t_embd) { | ||
| //cross.t_embd = t_embd; | ||
|
|
@@ -1005,11 +1015,8 @@ int llama_context::decode(const llama_batch & batch_inp) { | |
| n_outputs = n_outputs_new; | ||
| } | ||
|
|
||
| ggml_status status; | ||
| const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status); | ||
|
|
||
| if (!res) { | ||
| // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache | ||
|
|
@@ -1190,10 +1197,6 @@ int llama_context::decode(const llama_batch & batch_inp) { | |
| // wait for the computation to finish (automatically done when obtaining the model output) | ||
| //synchronize(); | ||
|
|
||
| return 0; | ||
| } | ||
|
|
||
|
|
@@ -1275,20 +1278,8 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) { | |
| // graph | ||
| // | ||
|
|
||
| uint32_t llama_context::graph_max_nodes() const { | ||
| return std::max<uint32_t>(65536u, 5u*model.n_tensors()); | ||
|
There was a problem hiding this comment. On a side note, I don't know why the minimum number of nodes is so big. This is a waste of memory for the smaller models. The change seems to be from #11571. There was a problem hiding this comment. I will follow-up with a PR to fix this. |
||
| } | ||
|
|
||
| ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx) { | ||
|
|
@@ -1301,6 +1292,9 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u | |
| LLAMA_LOG_DEBUG("%s: making n_tokens a multiple of n_seqs - n_tokens = %u, n_seqs = %u, n_outputs = %u\n", __func__, n_tokens, n_seqs, n_outputs); | ||
| } | ||
|
|
||
| gf_res_prev->reset(); | ||
| ggml_backend_sched_reset(sched.get()); | ||
|
|
||
| // store the n_outputs as it is, and restore it afterwards | ||
| // TODO: not sure if needed, might simplify in the future by removing this | ||
| const auto save_n_outputs = this->n_outputs; | ||
|
|
@@ -1310,17 +1304,15 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u | |
| llama_batch_allocr balloc(model.hparams.n_pos_per_embd()); | ||
| llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs); | ||
|
|
||
| auto * res = gf_res_reserve.get(); | ||
|
|
||
| const auto gparams = graph_params(res, ubatch, mctx, LLM_GRAPH_TYPE_DEFAULT); | ||
|
|
||
| res->reset(); | ||
|
|
||
| auto * gf = model.build_graph(gparams); | ||
|
|
||
| this->n_outputs = save_n_outputs; | ||
|
|
||
| // initialize scheduler with the specified graph | ||
| if (!ggml_backend_sched_reserve(sched.get(), gf)) { | ||
|
|
@@ -1331,28 +1323,27 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u | |
| return gf; | ||
| } | ||
|
|
||
| llm_graph_params llama_context::graph_params( | ||
| llm_graph_result_i * res, | ||
| const llama_ubatch & ubatch, | ||
| const llama_memory_context_i * mctx, | ||
| llm_graph_type gtype) const { | ||
| return { | ||
| /*.arch =*/ model.arch, | ||
| /*.hparams =*/ model.hparams, | ||
| /*.cparams =*/ cparams, | ||
| /*.ubatch =*/ ubatch, | ||
| /*.gtype =*/ gtype, | ||
| /*.sched =*/ sched.get(), | ||
| /*.backend_cpu =*/ backend_cpu, | ||
| /*.cvec =*/ &cvec, | ||
| /*.loras =*/ &loras, | ||
| /*.mctx =*/ mctx, | ||
| /*.cross =*/ &cross, | ||
| /*.n_outputs =*/ n_outputs, | ||
| /*.cb =*/ graph_get_cb(), | ||
| /*.res =*/ res, | ||
| }; | ||
| } | ||
|
|
||
| ggml_status llama_context::graph_compute( | ||
|
|
@@ -1930,6 +1921,7 @@ llama_perf_context_data llama_context::perf_get_data() const { | |
| data.t_eval_ms = 1e-3 * t_eval_us; | ||
| data.n_p_eval = std::max(1, n_p_eval); | ||
| data.n_eval = std::max(1, n_eval); | ||
| data.n_reused = std::max(0, n_reused); | ||
|
|
||
| return data; | ||
| } | ||
|
|
@@ -1938,6 +1930,7 @@ void llama_context::perf_reset() { | |
| t_start_us = ggml_time_us(); | ||
| t_eval_us = n_eval = 0; | ||
| t_p_eval_us = n_p_eval = 0; | ||
| n_reused = 0; | ||
| } | ||
|
|
||
| // | ||
|
|
@@ -2064,8 +2057,13 @@ void llama_context::opt_epoch_iter( | |
| break; | ||
| } | ||
|
|
||
| auto * res = gf_res_prev.get(); | ||
|
|
||
| const auto gparams = graph_params(res, ubatch, mctx.get(), LLM_GRAPH_TYPE_DEFAULT); | ||
|
|
||
| res->reset(); | ||
|
|
||
| auto * gf = model.build_graph(gparams); | ||
|
|
||
| struct ggml_context * ctx_compute_opt; | ||
| { | ||
|
|
@@ -2807,6 +2805,7 @@ void llama_perf_context_print(const llama_context * ctx) { | |
| LLAMA_LOG_INFO("%s: eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", | ||
| __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval); | ||
| LLAMA_LOG_INFO("%s: total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval)); | ||
| LLAMA_LOG_INFO("%s: graphs reused = %10d\n", __func__, data.n_reused); | ||
| } | ||
|
|
||
| void llama_perf_context_reset(llama_context * ctx) { | ||
|
|
||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Uh oh!
There was an error while loading. Please reload this page.