From 9e5db0ce29b3b963c79bd03e691a264d0e6cd331 Mon Sep 17 00:00:00 2001 From: Finbarr Timbers Date: Mon, 21 Jul 2025 16:18:03 -0600 Subject: [PATCH 1/2] Fix misnamed variables. --- open_instruct/grpo_fast.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/open_instruct/grpo_fast.py b/open_instruct/grpo_fast.py index fe02a5bb29..b1999da3d9 100644 --- a/open_instruct/grpo_fast.py +++ b/open_instruct/grpo_fast.py @@ -1865,7 +1865,7 @@ def maybe_evaluate( try: # timeout 0.01 if this is the last training step or we're not evaluating # otherwise, wait to get the last evaluation generations (long timeout just in case) - timeout = 0.01 if (training_step < args.num_training_steps or args.eval_freq < 0) else 100 + timeout = 0.01 if (training_step < args.num_training_steps or args.local_eval_freq < 0) else 100 eval_result = evaluation_inference_results_Q.get(timeout=timeout) logger.info("[Main Thread] 📊 Evaluation responses received") @@ -2102,7 +2102,7 @@ def main(args: Args, tc: TokenizerConfig, model_config: ModelConfig, num_eval_sa # Start vLLM engines to process from queues for engine in vllm_engines: engine.process_from_queue.remote( - generation_config, eval_generation_config, args.eval_freq, args.num_training_steps, resume_training_step + generation_config, eval_generation_config, args.local_eval_freq, args.num_training_steps, resume_training_step ) logger.info("======== ✅ vllm engines started processing from queues =========") From 9b890dc7e761e0376df212d472d950240e1675f2 Mon Sep 17 00:00:00 2001 From: Finbarr Timbers Date: Mon, 21 Jul 2025 16:18:38 -0600 Subject: [PATCH 2/2] Ran linter. --- open_instruct/grpo_fast.py | 6 +++++- uv.lock | 2 +- 2 files changed, 6 insertions(+), 2 deletions(-) diff --git a/open_instruct/grpo_fast.py b/open_instruct/grpo_fast.py index b1999da3d9..7f644ec925 100644 --- a/open_instruct/grpo_fast.py +++ b/open_instruct/grpo_fast.py @@ -2102,7 +2102,11 @@ def main(args: Args, tc: TokenizerConfig, model_config: ModelConfig, num_eval_sa # Start vLLM engines to process from queues for engine in vllm_engines: engine.process_from_queue.remote( - generation_config, eval_generation_config, args.local_eval_freq, args.num_training_steps, resume_training_step + generation_config, + eval_generation_config, + args.local_eval_freq, + args.num_training_steps, + resume_training_step, ) logger.info("======== ✅ vllm engines started processing from queues =========") diff --git a/uv.lock b/uv.lock index d63769582c..bda1b2c521 100644 --- a/uv.lock +++ b/uv.lock @@ -2409,7 +2409,7 @@ requires-dist = [ { name = "debugpy", specifier = ">=1.8.13" }, { name = "deepspeed", specifier = "==0.15.4" }, { name = "fastapi", marker = "extra == 'code'", specifier = ">=0.100.0" }, - { name = "flash-attn", marker = "extra == 'compile'", specifier = ">=2.8.0.post1" }, + { name = "flash-attn", marker = "extra == 'compile'", specifier = ">=2.8.0.post2" }, { name = "flashinfer-python", marker = "extra == 'compile'", specifier = ">=0.2.7.post1" }, { name = "hf-transfer", specifier = ">=0.1.8" }, { name = "immutabledict", specifier = "==1.2.0" },