b300 dsv3 bf16 hang fix

malay-nagda · malay-nagda · commit f25426c1d201 · 2026-02-06T19:09:07.000+05:30
Signed-off-by: Malay Nagda &lt;malayn@nvidia.com&gt;
diff --git a/scripts/performance/configs/deepseek/deepseek_workload_base_configs.py b/scripts/performance/configs/deepseek/deepseek_workload_base_configs.py
@@ -149,6 +149,7 @@
 DEEPSEEK_V3_PRETRAIN_CONFIG_B300_V2 = replace(
     DEEPSEEK_V3_PRETRAIN_CONFIG_B300_V1,
     global_batch_size=4096,
+    recompute_modules=["mla_up_proj", "moe_act", "layernorm", "mlp", "moe"],
 )
 DEEPSEEK_V3_PRETRAIN_CONFIG_B300_BF16_V2 = replace(
     DEEPSEEK_V3_PRETRAIN_CONFIG_B300_V2,

Original file line number	Diff line number	Diff line change
`@@ -149,6 +149,7 @@`
`149`	`149`	`DEEPSEEK_V3_PRETRAIN_CONFIG_B300_V2 = replace(`
`150`	`150`	`DEEPSEEK_V3_PRETRAIN_CONFIG_B300_V1,`
`151`	`151`	`global_batch_size=4096,`
	`152`	`+ recompute_modules=["mla_up_proj", "moe_act", "layernorm", "mlp", "moe"],`
`152`	`153`	`)`
`153`	`154`	`DEEPSEEK_V3_PRETRAIN_CONFIG_B300_BF16_V2 = replace(`
`154`	`155`	`DEEPSEEK_V3_PRETRAIN_CONFIG_B300_V2,`