Update README.md
Browse files
README.md
CHANGED
|
@@ -114,12 +114,15 @@ The model was evaluated on AIME24, GPQA Diamond, MATH-500, and GSM8K benchmarks.
|
|
| 114 |
The results of AIME24, MATH-500 and GPQA Diamond were obtained using forked [lighteval](https://github.com/zhaolin-amd/lighteval/tree/v0.10-release-custom) and vLLM docker `rocm/vllm-private:pytorch-vllm-gfx950-mxfp4-mxfp6-v3`.
|
| 115 |
|
| 116 |
```
|
|
|
|
| 117 |
export VLLM_QUARK_F4F6_OFFLINE_DEQUANT_TMPENVVAR=1
|
| 118 |
|
|
|
|
| 119 |
MODEL_ARGS="model_name=amd/DeepSeek-R1-0528-MXFP4-ASQ,dtype=bfloat16,tensor_parallel_size=8,max_model_length=71536,max_num_batched_tokens=32768,gpu_memory_utilization=0.85,generation_parameters={max_new_tokens:65536,temperature:0.6,top_p:0.95,seed:$SEED}"
|
| 120 |
OUTPUT_DIR="results/DeepSeek-R1-0528-MXFP4-ASQ-Seed"
|
| 121 |
LOG="logs/deepseek_0528_maxfp4.log"
|
| 122 |
|
|
|
|
| 123 |
for i in $(seq 1 10); do
|
| 124 |
# seed in [0, 2**30 - 1]
|
| 125 |
SEED=$(shuf -i 0-1073741823 -n 1)
|
|
|
|
| 114 |
The results of AIME24, MATH-500 and GPQA Diamond were obtained using forked [lighteval](https://github.com/zhaolin-amd/lighteval/tree/v0.10-release-custom) and vLLM docker `rocm/vllm-private:pytorch-vllm-gfx950-mxfp4-mxfp6-v3`.
|
| 115 |
|
| 116 |
```
|
| 117 |
+
# Set docker env
|
| 118 |
export VLLM_QUARK_F4F6_OFFLINE_DEQUANT_TMPENVVAR=1
|
| 119 |
|
| 120 |
+
# Set model args
|
| 121 |
MODEL_ARGS="model_name=amd/DeepSeek-R1-0528-MXFP4-ASQ,dtype=bfloat16,tensor_parallel_size=8,max_model_length=71536,max_num_batched_tokens=32768,gpu_memory_utilization=0.85,generation_parameters={max_new_tokens:65536,temperature:0.6,top_p:0.95,seed:$SEED}"
|
| 122 |
OUTPUT_DIR="results/DeepSeek-R1-0528-MXFP4-ASQ-Seed"
|
| 123 |
LOG="logs/deepseek_0528_maxfp4.log"
|
| 124 |
|
| 125 |
+
# Evaluating 10 rounds
|
| 126 |
for i in $(seq 1 10); do
|
| 127 |
# seed in [0, 2**30 - 1]
|
| 128 |
SEED=$(shuf -i 0-1073741823 -n 1)
|