Relax GSM8K thresholds for the GLM-5.2 DSA-MTP variants (#38725)
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
354ee46c26
commit
95a88bfd69
@@ -24,6 +24,7 @@ class TestGLM52NVFP4DPMTP(
|
||||
mem_fraction_static = 0.88
|
||||
enable_dp_attention = True
|
||||
bs_1_speed_thres = 180
|
||||
gsm8k_accuracy_thres = 0.92
|
||||
extra_server_args = [
|
||||
"--moe-runner-backend",
|
||||
"flashinfer_trtllm",
|
||||
|
||||
@@ -22,6 +22,7 @@ class TestGLM52NVFP4TPMTP(
|
||||
tp_size = 4
|
||||
mem_fraction_static = 0.8
|
||||
bs_1_speed_thres = 280
|
||||
gsm8k_accuracy_thres = 0.92
|
||||
extra_server_args = [
|
||||
"--moe-runner-backend",
|
||||
"flashinfer_trtllm",
|
||||
|
||||
Reference in New Issue
Block a user