Add rl example (#49)

* update sampler

* update sampler

* updat

* update sampler

* update cpu env

* update compat

* update

* update

* update

* update

* update server

* fix

* fix

* fix smaple

* fix smaple

* update

* update

* update server

Author: Yunnglin

cookbook/client/tinker/transformer/grpo.py

@@ -34,14 +34,14 @@
 logger = get_logger()
 
 # ========== Configuration ==========
-BASE_MODEL = 'Qwen/Qwen2.5-0.5B-Instruct'
+BASE_MODEL = 'Qwen/Qwen2.5-3B-Instruct'
 NUM_GENERATIONS = 4
 MAX_NEW_TOKENS = 1024
 LEARNING_RATE = 1e-5
-MAX_STEPS = 10
-BATCH_SIZE = 1
+MAX_STEPS = 100
+BATCH_SIZE = 2
 TEMPERATURE = 1.0
-SYNC_INTERVAL = 5       # Save weights for sampler every N steps
+SYNC_INTERVAL = 2       # Save weights for sampler every N steps
 LORA_RANK = 8