From 8cbbda005cd79577731e31d3e1862f74bd8b68e1 Mon Sep 17 00:00:00 2001
From: RileyChen23 <2017009663@qq.com>
Date: Tue, 14 Jul 2026 18:41:40 +0800
Subject: [PATCH] fix: tighten runtime and GRPO metrics

---
 .gitignore            | 4 +++-
 requirements.txt      | 4 +++-
 trainer/train_grpo.py | 2 +-
 3 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/.gitignore b/.gitignore
index 7956e1d..0dca8ff 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,6 +1,8 @@
 __pycache__/
 *.pyc
+.pytest_cache/
+.codex_pytest_tmp*/
 .DS_Store
 out
 website/
-docs-minimind/
\ No newline at end of file
+docs-minimind/
diff --git a/requirements.txt b/requirements.txt
index 12790f4..99711f5 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,5 +1,6 @@
 datasets==3.6.0
 datasketch==1.6.4
+fastapi==0.115.6
 Flask==3.0.3
 Flask_Cors==4.0.0
 jieba==0.42.1
@@ -23,10 +24,11 @@ jinja2==3.1.2
 jsonlines==4.0.0
 trl==0.13.0
 ujson==5.1.0
+uvicorn==0.34.0
 wandb==0.18.3
 streamlit==1.50.0
 einops==0.8.1
 swanlab==0.7.11
 modelscope==1.37.0
 # torch==2.6.0
-# torchvision==0.21.0
\ No newline at end of file
+# torchvision==0.21.0
diff --git a/trainer/train_grpo.py b/trainer/train_grpo.py
index f4de0a7..b32f8e1 100755
--- a/trainer/train_grpo.py
+++ b/trainer/train_grpo.py
@@ -168,7 +168,7 @@ def grpo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, reward_mod
             current_aux_loss = aux_loss.item()
             avg_reward_val = rewards.mean().item()
             avg_len_val = completion_mask.sum(dim=1).float().mean().item()
-            kl_ref_val = ((ref_per_token_logps - per_token_logps) * completion_mask).sum().item() / max(completion_mask.sum().item(), 1)
+            kl_ref_val = (per_token_kl * completion_mask).sum().item() / max(completion_mask.sum().item(), 1)
             advantages_mean_val = advantages.mean().item()
             advantages_std_val = advantages.std().item()
             current_lr = optimizer.param_groups[0]['lr']