[DSv32] [GLM5] Improve Model Quality by Avoiding FP32 Precision Loss in weights_proj (#19041)

2026-02-22 16:20:51 +08:00
parent 326b788ab4
commit eddf193292
4 changed files with 48 additions and 9 deletions
@@ -24,7 +24,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMo
 from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
 from sglang.test.test_utils import CustomTestCase

-register_cuda_ci(est_time=2, suite="stage-b-test-small-1-gpu")
+register_cuda_ci(est_time=2, suite="stage-b-test-large-1-gpu")

 # Global configuration for all indexer tests
 DEFAULT_CONFIG = {
@@ -34,7 +34,7 @@ DEFAULT_CONFIG = {
    "context_len": 2048,
    "max_bs": 64,
    "hidden_size": 5120,
-    "index_n_heads": 1,
+    "index_n_heads": 32,
    "index_head_dim": 128,
    "rope_head_dim": 64,
    "index_topk": 64,