更新调度策略

ztxz16 · Aug 14, 2024 · 5c282eb · 5c282eb
1 parent 4d1b144
commit 5c282eb
Showing 1 changed file with 31 additions and 9 deletions.
diff --git a/src/models/basellm.cpp b/src/models/basellm.cpp
@@ -526,10 +526,10 @@ namespace fastllm {
 
                     model->tokensLimit = maxTotalLens;
                     int limit = maxTotalLens;
-                    model->promptLimit = limit * 2 / 3;
+                    model->promptLimit = limit * 3 / 4;
 
                     if (model->verbose) {
-                        printf("Fastllm KV Cache Limit: %f MB.\n", (double)kvCacheLimit / 1024 / 1024);
+                        printf("Fastllm KV Cache Limit: %f MB.\n", (double)kvCacheLimit / 1e6);
                         printf("Fastllm KV Cache Token limit: %d tokens.\n", maxTotalLens);
                         printf("Fastllm Prompt Token limit: %d tokens.\n", std::min(model->max_positions, model->promptLimit));
                         printf("Fastllm Batch limit: %d.\n", maxBatch);
@@ -567,10 +567,11 @@ namespace fastllm {
                         int limit = maxTotalLens;
                         int promptLimit = model->promptLimit;
 
-                        int lenSum = 0;
+                        int lenSum = 0, currentActivate = 0;
                         for (auto &it: model->responseContextDict.dicts) {
                             if (it.second->pastKeyValues[0].first.expansionDims.size() > 0) {
                                 lenSum += it.second->pastKeyValues[0].first.expansionDims[1];
+                                currentActivate++;
                             }
                         }
 
@@ -579,10 +580,11 @@ namespace fastllm {
                             if (isPrompt == 0 && seqLens.size() > 0) {
                                 continue;
                             }
+/*
                             if (lenSum >= promptLimit && isPrompt) {
                                 continue;
                             }
-
+*/
                             for (auto &it: model->responseContextDict.dicts) {
                                 if (it.second->isEnding) {
                                     continue;
@@ -593,17 +595,23 @@ namespace fastllm {
                                 if (!isPrompt && it.second->preTokens == 0) {
                                     continue;
                                 }
-                                if (it.second->currentTokens.size() > promptLimit) {
+
+                                if (it.second->currentTokens.size() > maxTotalLens) {
                                     it.second->isEnding = true;
                                     it.second->error = ResponseContextErrorPromptTooLong;
                                     continue;
                                 }
 
                                 int outputLimit = it.second->generationConfig.output_token_limit;
                                 outputLimit = (outputLimit < 0 ? 128 : outputLimit);
+/*
                                 if (isPrompt && lenSum + it.second->currentTokens.size() > promptLimit) {
                                     continue;
                                 }
+*/
+                                if (isPrompt && lenSum + it.second->currentTokens.size() + (currentActivate + 1) * 256 > maxTotalLens) {
+                                    continue;
+                                }
 
                                 if (!isPrompt) {
                                     if (it.second->pastKeyValues[0].first.expansionDims[1] == it.second->pastKeyValues[0].first.dims[1]) {
@@ -617,6 +625,9 @@ namespace fastllm {
                                         }
                                         lenSum += predictLen;
                                     }
+                                } else {
+                                    lenSum += it.second->currentTokens.size();
+                                    currentActivate++;
                                 }
 
                                 generationConfigs.push_back(it.second->generationConfig);
@@ -720,26 +731,37 @@ auto st = std::chrono::system_clock::now();
                                                                         *pastKeyValue1, generationConfigs[0], tokensManager, logits[0])};
                                 }
                             }
-
-
 //PrintProfiler();
 /*int total = 0;
 for (int i : seqLens) total += i;
 float spend = GetSpan(st, std::chrono::system_clock::now());
 printf("len = %d, spend = %f s. tokens / s = %f\n", (int)total, spend, (float)total / spend);
 */
+                            dictLocker.lock();
+
                             if (model->verbose) {
                                 genTokens += seqLens.size();
                                 auto nowTime = std::chrono::system_clock::now();
                                 float spend = GetSpan(lastRecordTime, nowTime);
                                 if (spend > 1) {
-                                    printf("Current batch: %d, Speed: %f tokens / s.\n", (int)seqLens.size(), (float)genTokens / spend);
+                                    int total = 0, alive = 0, aliveLen = 0, pending = 0;
+                                    for (auto &it: model->responseContextDict.dicts) {
+                                        if (it.second->isEnding) {
+                                            continue;
+                                        }
+                                        if (it.second->pastKeyValues[0].first.expansionDims.size() > 0) {
+                                            alive++;
+                                            aliveLen += it.second->pastKeyValues[0].first.expansionDims[1];
+                                        } else {
+                                            pending++;
+                                        }
+                                    }
+                                    printf("alive = %d, pending = %d, contextLen = %d, Speed: %f tokens / s.\n", alive, pending, aliveLen, (int)seqLens.size(), (float)genTokens / spend);
                                     lastRecordTime = nowTime;
                                     genTokens = 0;
                                 }
                             }
 
-                            dictLocker.lock();
                             for (int i = 0; i < handles.size(); i++) {
                                 auto &it = *model->responseContextDict.dicts.find(handles[i]);
                                 int curRet = ret[i];