Update H200 ultra-tier settings
Browse files- code/train_sft.py +2 -2
code/train_sft.py
CHANGED
|
@@ -169,8 +169,8 @@ def auto_profile_hardware():
|
|
| 169 |
# Tier Classification for 2048 Context Length (Target Effective Batch = 128)
|
| 170 |
if vram_gb >= 75: # H200 (141GB), H100 (80GB), GH200, A100-80GB
|
| 171 |
tier = "Ultra-Tier"
|
| 172 |
-
micro_batch =
|
| 173 |
-
grad_accum =
|
| 174 |
desc = "NVIDIA Hopper / Datacenter Beast (141GB / 80GB HBM3e)"
|
| 175 |
elif vram_gb >= 30: # RTX 5090 (32GB), A100 (40GB), A6000 (48GB), RTX 6000 Ada
|
| 176 |
tier = "High-Tier"
|
|
|
|
| 169 |
# Tier Classification for 2048 Context Length (Target Effective Batch = 128)
|
| 170 |
if vram_gb >= 75: # H200 (141GB), H100 (80GB), GH200, A100-80GB
|
| 171 |
tier = "Ultra-Tier"
|
| 172 |
+
micro_batch = 16
|
| 173 |
+
grad_accum = 8
|
| 174 |
desc = "NVIDIA Hopper / Datacenter Beast (141GB / 80GB HBM3e)"
|
| 175 |
elif vram_gb >= 30: # RTX 5090 (32GB), A100 (40GB), A6000 (48GB), RTX 6000 Ada
|
| 176 |
tier = "High-Tier"
|