Set expandable_segments in the 2 x DGX Spark environment
#8
by Bizuayeu - opened
README.md
CHANGED
|
@@ -213,6 +213,7 @@ export VLLM_ENABLE_PCIE_ALLREDUCE="0"
|
|
| 213 |
export B12X_POLICY_MODE="auto"
|
| 214 |
export VLLM_B12X_MOE_FP4_FORCE_A16=1
|
| 215 |
export B12X_W4A16_A4_PREFILL=1
|
|
|
|
| 216 |
```
|
| 217 |
|
| 218 |
Launch the head node:
|
|
|
|
| 213 |
export B12X_POLICY_MODE="auto"
|
| 214 |
export VLLM_B12X_MOE_FP4_FORCE_A16=1
|
| 215 |
export B12X_W4A16_A4_PREFILL=1
|
| 216 |
+
export PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True"
|
| 217 |
```
|
| 218 |
|
| 219 |
Launch the head node:
|