Class DecodeCudaGraphLog
java.lang.Object
smile.llm.engine.DecodeCudaGraphLog
INFO logging for decode CUDA graph bucket lifecycle (once per bucket event).
Used to compare on-demand capture hitches vs pre-captured bucket promotion.
-
Method Summary
Modifier and TypeMethodDescriptionstatic voidbucketCapture(int tpRank, int batch, int numPages, long captureMs, boolean prefetched) Logs successful graph capture for a bucket.static voidbucketReplay(int tpRank, int batch, int numPages) Logs the first replay for a bucket in this process.static voidbucketWarmup(int tpRank, int batch, int numPages, int step, int total) Logs the first eager warmup step for a bucket.static voidprefetchHit(int tpRank, int batch, int numPages) Logs promotion of a prefetched graph at a page boundary (no capture hitch).static voidprefetchReady(int tpRank, int batch, int numPages, long captureMs) Logs when a prefetched bucket becomes replay-ready.static voidprefetchStart(int tpRank, int batch, int numPages) Logs when background prefetch starts for the next page bucket.static voidprefetchWarmup(int tpRank, int batch, int numPages, int step, int total) Logs the first eager warmup step during next-bucket prefetch capture.
-
Method Details
-
bucketWarmup
public static void bucketWarmup(int tpRank, int batch, int numPages, int step, int total) Logs the first eager warmup step for a bucket.- Parameters:
tpRank- tensor-parallel rank for logging.batch- decode batch size.numPages- KV page count for this bucket.step- current warmup step (1-based).total- total warmup steps before capture.
-
bucketCapture
public static void bucketCapture(int tpRank, int batch, int numPages, long captureMs, boolean prefetched) Logs successful graph capture for a bucket.- Parameters:
tpRank- tensor-parallel rank for logging.batch- decode batch size.numPages- KV page count for this bucket.captureMs- capture duration in milliseconds.prefetched-truewhen capture was a next-bucket prefetch.
-
bucketReplay
public static void bucketReplay(int tpRank, int batch, int numPages) Logs the first replay for a bucket in this process.- Parameters:
tpRank- tensor-parallel rank for logging.batch- decode batch size.numPages- KV page count for this bucket.
-
prefetchWarmup
public static void prefetchWarmup(int tpRank, int batch, int numPages, int step, int total) Logs the first eager warmup step during next-bucket prefetch capture.- Parameters:
tpRank- tensor-parallel rank for logging.batch- decode batch size.numPages- KV page count for this bucket.step- current warmup step (1-based).total- total warmup steps before capture.
-
prefetchStart
public static void prefetchStart(int tpRank, int batch, int numPages) Logs when background prefetch starts for the next page bucket.- Parameters:
tpRank- tensor-parallel rank for logging.batch- decode batch size.numPages- KV page count for this bucket.
-
prefetchReady
public static void prefetchReady(int tpRank, int batch, int numPages, long captureMs) Logs when a prefetched bucket becomes replay-ready.- Parameters:
tpRank- tensor-parallel rank for logging.batch- decode batch size.numPages- KV page count for this bucket.captureMs- capture duration in milliseconds.
-
prefetchHit
public static void prefetchHit(int tpRank, int batch, int numPages) Logs promotion of a prefetched graph at a page boundary (no capture hitch).- Parameters:
tpRank- tensor-parallel rank for logging.batch- decode batch size.numPages- KV page count for this bucket.
-