Remove fp8 kv cache option from grpo_fast (#1203)

Author: finbarrtimbers

open_instruct/benchmark_generators.py

@@ -263,7 +263,6 @@ def setup_vllm_engines(
         prompt_queue=param_prompt_Q,
         results_queue=inference_results_Q,
         actor_manager=actor_manager,
-        use_fp8_kv_cache=args.use_fp8_kv_cache,
         inflight_updates=args.inflight_updates,
     )
 

open_instruct/grpo_fast.py

@@ -227,9 +227,6 @@ class Args:
     stop_strings: list[str] | None = None
     """List of strings that stop the generation when they are generated.
     The returned output will not contain the stop strings."""
-    use_fp8_kv_cache: bool = False
-    """Whether to use fp8 kv cache. This is useful for larger models or olmo."""
-
     # Algorithm
     async_steps: int = 1
     """Number of steps ahead to generate responses. Set to 0 to make the code synchronous. Values greater than 0 learn from a policy up to async_steps old like Cleanba (https://arxiv.org/abs/2310.00036)"""
@@ -2336,7 +2333,6 @@ def create_model_and_optimizer(
         results_queue=inference_results_Q,
         eval_results_queue=evaluation_inference_results_Q,
         actor_manager=actor_manager,
-        use_fp8_kv_cache=args.use_fp8_kv_cache,
         inflight_updates=args.inflight_updates,
     )
 

scripts/train/olmo3/32b_rl_smoke_test.sh

@@ -69,7 +69,6 @@ python mason.py \
         --llm_judge_max_context_length 32768 \
         --clip_higher 0.272 \
         --allow_world_padding False \
-        --use_fp8_kv_cache False \
         --code_api_url https://p9f1719l7f.execute-api.us-west-2.amazonaws.com/prod/test_program \
         --code_pass_rate_reward_threshold 0.99 \
         --oe_eval_max_length 32768 \