llama : llama_perf + option to disable timings during decode (ggml-org#9355)

ggerganov · ngxson · arthw · commit 308df3bace5b · 2024-11-15T12:34:22.000+08:00
* llama : llama_perf + option to disable timings during decode

ggml-ci

* common : add llama_arg

* Update src/llama.cpp

Co-authored-by: Xuan Son Nguyen &lt;thichthat@gmail.com&gt;

* perf : separate functions in the API

ggml-ci

* perf : safer pointer handling + naming update

ggml-ci

* minor : better local var name

* perf : abort on invalid sampler pointer

ggml-ci

---------

Co-authored-by: Xuan Son Nguyen &lt;thichthat@gmail.com&gt;
diff --git a/common/arg.cpp b/common/arg.cpp
@@ -720,6 +720,14 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
             params.prompt = value;
         }
     ));
+    add_opt(llama_arg(
+        {"--no-perf"},
+        format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
+        [](gpt_params & params) {
+            params.no_perf = true;
+            params.sparams.no_perf = true;
+        }
+    ).set_env("LLAMA_ARG_NO_PERF"));
     add_opt(llama_arg(
         {"-f", "--file"}, "FNAME",
         "a file containing the prompt (default: none)",
diff --git a/common/common.cpp b/common/common.cpp
@@ -820,7 +820,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         }
         llama_kv_cache_clear(lctx);
         llama_synchronize(lctx);
-        llama_perf_reset(lctx, LLAMA_PERF_TYPE_CONTEXT);
+        llama_perf_context_reset(lctx);
     }
 
     iparams.model   = model;
@@ -916,6 +916,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
     cparams.flash_attn        = params.flash_attn;
+    cparams.no_perf           = params.no_perf;
 
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
     cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
diff --git a/common/common.h b/common/common.h
@@ -124,6 +124,7 @@ struct gpt_sampler_params {
     float   mirostat_eta      = 0.10f; // learning rate
     bool    penalize_nl       = false; // consider newlines as a repeatable token
     bool    ignore_eos        = false;
+    bool    no_perf           = false; // disable performance metrics
 
     std::vector<enum gpt_sampler_type> samplers = {
         GPT_SAMPLER_TYPE_TOP_K,
@@ -246,6 +247,7 @@ struct gpt_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool flash_attn        = false; // flash attention
+    bool no_perf           = false; // disable performance metrics
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool logits_all        = false; // return logits for all tokens in the batch
diff --git a/common/sampling.cpp b/common/sampling.cpp
@@ -142,7 +142,7 @@ std::string gpt_sampler_params::print() const {
 struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params) {
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
-    lparams.no_perf = false; // TODO: control via params
+    lparams.no_perf = params.no_perf;
 
     auto * result = new gpt_sampler {
         /* .params = */ params,
@@ -257,10 +257,10 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *
     // TODO: measure grammar performance
 
     if (gsmpl) {
-        llama_perf_print(gsmpl->chain, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
+        llama_perf_sampler_print(gsmpl->chain);
     }
     if (ctx) {
-        llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+        llama_perf_context_print(ctx);
     }
 }
 
diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp
@@ -187,7 +187,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     llama_batch_free(batch);
 
diff --git a/examples/batched.swift/Sources/main.swift b/examples/batched.swift/Sources/main.swift
@@ -200,8 +200,8 @@ let t_main_end = ggml_time_us()
 
 print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n\n")
 
-llama_perf_print(UnsafeRawPointer(context), LLAMA_PERF_TYPE_CONTEXT)
-llama_perf_print(UnsafeRawPointer(smpl),    LLAMA_PERF_TYPE_SAMPLER_CHAIN)
+llama_perf_sampler_print(smpl)
+llama_perf_context_print(context)
 
 private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
     let utf8Count = text.utf8.count
diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp
@@ -229,8 +229,8 @@ int main(int argc, char ** argv) {
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
     LOG_TEE("\n");
-    llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
-    llama_perf_print(ctx,  LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_sampler_print(smpl);
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 
diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp
@@ -306,7 +306,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     // clean up
     llama_batch_free(batch);
diff --git a/examples/eval-callback/eval-callback.cpp b/examples/eval-callback/eval-callback.cpp
@@ -182,7 +182,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     llama_free(ctx);
     llama_free_model(model);
diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp
@@ -637,7 +637,7 @@ int main(int argc, char ** argv) {
     g_collector.save_imatrix();
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     llama_free(ctx);
     llama_free_model(model);
diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp
@@ -1630,7 +1630,7 @@ int main(int argc, char ** argv) {
             fflush(p_err->fout);
         }
 
-        llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+        llama_perf_context_print(ctx);
 
         llama_free(ctx);
 
diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp
@@ -308,7 +308,7 @@ int main(int argc, char ** argv) {
         // process the prompt
         process_prompt(ctx_llava, image_embed, &params, params.prompt);
 
-        llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
+        llama_perf_context_print(ctx_llava->ctx_llama);
         llava_image_embed_free(image_embed);
         ctx_llava->model = NULL;
         llava_free(ctx_llava);
@@ -325,7 +325,7 @@ int main(int argc, char ** argv) {
             // process the prompt
             process_prompt(ctx_llava, image_embed, &params, params.prompt);
 
-            llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
+            llama_perf_context_print(ctx_llava->ctx_llama);
             llava_image_embed_free(image_embed);
             ctx_llava->model = NULL;
             llava_free(ctx_llava);
diff --git a/examples/llava/minicpmv-cli.cpp b/examples/llava/minicpmv-cli.cpp
@@ -319,7 +319,7 @@ int main(int argc, char ** argv) {
             }
         }
         printf("\n");
-        llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
+        llama_perf_context_print(ctx_llava->ctx_llama);
 
         ctx_llava->model = NULL;
         llava_free(ctx_llava);
diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp
@@ -240,8 +240,7 @@ int main(int argc, char ** argv){
     LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
     LOG_TEE("\ntarget:\n\n");
-    llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
-    llama_perf_print(ctx,  LLAMA_PERF_TYPE_CONTEXT);
+    gpt_perf_print(ctx, smpl);
 
     gpt_sampler_free(smpl);
 
diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp
@@ -415,7 +415,7 @@ int main(int argc, char ** argv) {
     LOG_TEE("\n");
 
     // TODO: print sampling/grammar timings for all clients
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     llama_batch_free(batch);
 
diff --git a/examples/passkey/passkey.cpp b/examples/passkey/passkey.cpp
@@ -256,7 +256,7 @@ int main(int argc, char ** argv) {
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 
diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
@@ -2047,7 +2047,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
     write_logfile(ctx, params, model, results);
 
     llama_free(ctx);
diff --git a/examples/retrieval/retrieval.cpp b/examples/retrieval/retrieval.cpp
@@ -292,7 +292,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx);
 
     // clean up
     llama_batch_free(query_batch);
diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp
@@ -154,8 +154,8 @@ int main(int argc, char ** argv) {
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
     LOG_TEE("\n");
-    llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
-    llama_perf_print(ctx,  LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_sampler_print(smpl);
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
@@ -616,7 +616,7 @@ int main(int argc, char ** argv) {
 
     LOG_TEE("\ndraft:\n\n");
     // TODO: print sampling/grammar timings for all drafts
-    llama_perf_print(ctx_dft, LLAMA_PERF_TYPE_CONTEXT);
+    llama_perf_context_print(ctx_dft);
 
     LOG_TEE("\ntarget:\n\n");
     gpt_perf_print(ctx_tgt, smpl);
diff --git a/include/llama.h b/include/llama.h
@@ -343,7 +343,7 @@ extern "C" {
         bool embeddings;  // if true, extract embeddings (together with logits)
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
         bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
-      //bool no_perf;     // whether to measure performance timings, TODO: implement
+        bool no_perf;     // whether to measure performance timings
 
         // Abort callback
         // if it returns true, execution of llama_decode() will be aborted
@@ -1176,13 +1176,30 @@ extern "C" {
     // NOTE: Used by llama.cpp examples, avoid using in third-party apps. Instead, do your own performance measurements.
     //
 
-    enum llama_perf_type {
-        LLAMA_PERF_TYPE_CONTEXT       = 0,
-        LLAMA_PERF_TYPE_SAMPLER_CHAIN = 1,
+    struct llama_perf_context_data {
+        double t_start_ms;
+        double t_load_ms;
+        double t_p_eval_ms;
+        double t_eval_ms;
+
+        int32_t n_p_eval;
+        int32_t n_eval;
     };
 
-    LLAMA_API void llama_perf_print(const void * ctx, enum llama_perf_type type);
-    LLAMA_API void llama_perf_reset(      void * ctx, enum llama_perf_type type);
+    struct llama_perf_sampler_data {
+        double t_sample_ms;
+
+        int32_t n_sample;
+    };
+
+    LLAMA_API struct llama_perf_context_data llama_perf_context      (const struct llama_context * ctx);
+    LLAMA_API void                           llama_perf_context_print(const struct llama_context * ctx);
+    LLAMA_API void                           llama_perf_context_reset(      struct llama_context * ctx);
+
+    // NOTE: the following work only with samplers constructed via llama_sampler_chain_init
+    LLAMA_API struct llama_perf_sampler_data llama_perf_sampler      (const struct llama_sampler * chain);
+    LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
+    LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
     LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx);
 
diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp
@@ -1669,3 +1669,37 @@ uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl) {
 
     return LLAMA_DEFAULT_SEED;
 }
+
+// perf
+
+struct llama_perf_sampler_data llama_perf_sampler(const struct llama_sampler * chain) {
+    struct llama_perf_sampler_data data = {};
+
+    if (chain == nullptr || chain->iface != &llama_sampler_chain_i) {
+        GGML_ABORT("%s: invalid sampler passed - requires a sampler created with llama_sampler_chain_init()\n", __func__);
+    }
+
+    const auto * ctx = (const struct llama_sampler_chain *) chain->ctx;
+
+    data.t_sample_ms = 1e-3 * ctx->t_sample_us;
+    data.n_sample    = std::max(0, ctx->n_sample);
+
+    return data;
+}
+
+void llama_perf_sampler_print(const struct llama_sampler * chain) {
+    const auto data = llama_perf_sampler(chain);
+
+    LLAMA_LOG_INFO("%s:    sampling time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
+            __func__, data.t_sample_ms, data.n_sample, data.t_sample_ms / data.n_sample, 1e3 / data.t_sample_ms * data.n_sample);
+}
+
+void llama_perf_sampler_reset(struct llama_sampler * chain) {
+    if (chain == nullptr || chain->iface != &llama_sampler_chain_i) {
+        GGML_ABORT("%s: invalid sampler passed - requires a sampler created with llama_sampler_chain_init()\n", __func__);
+    }
+
+    auto * ctx = (struct llama_sampler_chain *) chain->ctx;
+
+    ctx->t_sample_us = ctx->n_sample = 0;
+}
diff --git a/src/llama.cpp b/src/llama.cpp

Original file line number	Diff line number	Diff line change
`@@ -820,7 +820,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {`
`820`	`820`	`}`
`821`	`821`	`llama_kv_cache_clear(lctx);`
`822`	`822`	`llama_synchronize(lctx);`
`823`		`- llama_perf_reset(lctx, LLAMA_PERF_TYPE_CONTEXT);`
	`823`	`+ llama_perf_context_reset(lctx);`
`824`	`824`	`}`
`825`	`825`
`826`	`826`	`iparams.model = model;`
`@@ -916,6 +916,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param`
`916`	`916`	`cparams.cb_eval_user_data = params.cb_eval_user_data;`
`917`	`917`	`cparams.offload_kqv = !params.no_kv_offload;`
`918`	`918`	`cparams.flash_attn = params.flash_attn;`
	`919`	`+ cparams.no_perf = params.no_perf;`
`919`	`920`
`920`	`921`	`cparams.type_k = kv_cache_type_from_str(params.cache_type_k);`
`921`	`922`	`cparams.type_v = kv_cache_type_from_str(params.cache_type_v);`
Original file line number	Diff line number	Diff line change
`@@ -142,7 +142,7 @@ std::string gpt_sampler_params::print() const {`
`142`	`142`	`struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params) {`
`143`	`143`	`llama_sampler_chain_params lparams = llama_sampler_chain_default_params();`
`144`	`144`
`145`		`- lparams.no_perf = false; // TODO: control via params`
	`145`	`+ lparams.no_perf = params.no_perf;`
`146`	`146`
`147`	`147`	`auto * result = new gpt_sampler {`
`148`	`148`	`/* .params = */ params,`
`@@ -257,10 +257,10 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *`
`257`	`257`	`// TODO: measure grammar performance`
`258`	`258`
`259`	`259`	`if (gsmpl) {`
`260`		`- llama_perf_print(gsmpl->chain, LLAMA_PERF_TYPE_SAMPLER_CHAIN);`
	`260`	`+ llama_perf_sampler_print(gsmpl->chain);`
`261`	`261`	`}`
`262`	`262`	`if (ctx) {`
`263`		`- llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);`
	`263`	`+ llama_perf_context_print(ctx);`
`264`	`264`	`}`
`265`	`265`	`}`
`266`	`266`
Original file line number	Diff line number	Diff line change
`@@ -187,7 +187,7 @@ int main(int argc, char ** argv) {`
`187`	`187`	`}`
`188`	`188`
`189`	`189`	`LOG_TEE("\n");`
`190`		`- llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);`
	`190`	`+ llama_perf_context_print(ctx);`
`191`	`191`
`192`	`192`	`llama_batch_free(batch);`
`193`	`193`
Original file line number	Diff line number	Diff line change
`@@ -306,7 +306,7 @@ int main(int argc, char ** argv) {`
`306`	`306`	`}`
`307`	`307`
`308`	`308`	`LOG_TEE("\n");`
`309`		`- llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);`
	`309`	`+ llama_perf_context_print(ctx);`
`310`	`310`
`311`	`311`	`// clean up`
`312`	`312`	`llama_batch_free(batch);`
Original file line number	Diff line number	Diff line change
`@@ -182,7 +182,7 @@ int main(int argc, char ** argv) {`
`182`	`182`	`}`
`183`	`183`
`184`	`184`	`LOG_TEE("\n");`
`185`		`- llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);`
	`185`	`+ llama_perf_context_print(ctx);`
`186`	`186`
`187`	`187`	`llama_free(ctx);`
`188`	`188`	`llama_free_model(model);`