Tune prefill batching across Qwen profiles
This commit is contained in:
+8
-8
@@ -105,9 +105,9 @@ services:
|
||||
- --threads-batch
|
||||
- "${LLAMA_THREADS_BATCH:-6}"
|
||||
- --batch-size
|
||||
- "64"
|
||||
- "${FAST_BATCH_SIZE:-64}"
|
||||
- --ubatch-size
|
||||
- "32"
|
||||
- "${FAST_UBATCH_SIZE:-32}"
|
||||
- --parallel
|
||||
- "1"
|
||||
- --jinja
|
||||
@@ -268,9 +268,9 @@ services:
|
||||
- --threads-batch
|
||||
- "${LLAMA_THREADS_BATCH:-6}"
|
||||
- --batch-size
|
||||
- "64"
|
||||
- "${LARGE_BATCH_SIZE:-2048}"
|
||||
- --ubatch-size
|
||||
- "32"
|
||||
- "${LARGE_UBATCH_SIZE:-128}"
|
||||
- --parallel
|
||||
- "1"
|
||||
- --jinja
|
||||
@@ -345,9 +345,9 @@ services:
|
||||
- --threads-batch
|
||||
- "${LLAMA_THREADS_BATCH:-6}"
|
||||
- --batch-size
|
||||
- "64"
|
||||
- "${ULTRA_BATCH_SIZE:-2048}"
|
||||
- --ubatch-size
|
||||
- "32"
|
||||
- "${ULTRA_UBATCH_SIZE:-128}"
|
||||
- --parallel
|
||||
- "1"
|
||||
- --jinja
|
||||
@@ -426,9 +426,9 @@ services:
|
||||
- --threads-batch
|
||||
- "${LLAMA_THREADS_BATCH:-6}"
|
||||
- --batch-size
|
||||
- "64"
|
||||
- "${UNCENSORED_BATCH_SIZE:-2048}"
|
||||
- --ubatch-size
|
||||
- "32"
|
||||
- "${UNCENSORED_UBATCH_SIZE:-128}"
|
||||
- --parallel
|
||||
- "1"
|
||||
- --jinja
|
||||
|
||||
Reference in New Issue
Block a user