From d2441e030c791ad484f25425da43e22d0fcb762f Mon Sep 17 00:00:00 2001
From: mr-lin314 <798948055@qq.com>
Date: Fri, 29 Aug 2025 15:22:01 +0800
Subject: [PATCH 1/4] [docs]add mindspeed_rl commit id

---
 examples/rl/README.md | 1 +
 1 file changed, 1 insertion(+)

diff --git a/examples/rl/README.md b/examples/rl/README.md
index 27443a61..592b4f4b 100644
--- a/examples/rl/README.md
+++ b/examples/rl/README.md
@@ -92,6 +92,7 @@ cd ..
 # 下载MindSpeed RL
 git clone https://gitee.com/ascend/MindSpeed-RL.git
 cd MindSpeed-RL
+git checkout 6cb2e0d61eed19460d24575f65005930b8335329
 pip install -r requirements.txt
 cp -r mindspeed_rl ../MindSpeed-MM/
 cd ..
-- 
Gitee


From f15628caa1d5933fed6139d2c7c145419f97d71c Mon Sep 17 00:00:00 2001
From: mr-lin314 <798948055@qq.com>
Date: Fri, 29 Aug 2025 16:20:51 +0800
Subject: [PATCH 2/4] [docs]add mindspeed_rl commit id

---
 examples/rl/README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/examples/rl/README.md b/examples/rl/README.md
index 592b4f4b..e931d9e3 100644
--- a/examples/rl/README.md
+++ b/examples/rl/README.md
@@ -85,7 +85,6 @@ git clone https://gitee.com/ascend/MindSpeed.git
 cd MindSpeed
 # checkout commit from MindSpeed core_r0.12.1
 git checkout 6d63944cb2470a0bebc38dfb65299b91329b8d92
-pip install -r requirements.txt
 cp -r mindspeed ../MindSpeed-MM/
 cd ..
 
@@ -135,6 +134,7 @@ pip install numpy==1.26.4
 source /usr/local/Ascend/ascend-toolkit/set_env.sh
 source /usr/local/Ascend/nnal/atb/set_env.sh
 cd vllm-ascend
+pip install -r requirements-dev.txt
 # 因为前面已经安装了对应的torch_npu版本，这里需要将vllm_ascend文件中的requirements.txt中的torch-npu==2.5.1注释
 # 安装时若编译失败，可以关闭编译：export COMPILE_CUSTOM_KERNELS=0
 python setup.py develop
-- 
Gitee


From 07a0bea35eebbae59e3626f02bbb6dbeae8196c2 Mon Sep 17 00:00:00 2001
From: mr-lin314 <798948055@qq.com>
Date: Fri, 29 Aug 2025 17:00:26 +0800
Subject: [PATCH 3/4] [docs]add mindspeed_rl commit id

---
 examples/rl/README.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/examples/rl/README.md b/examples/rl/README.md
index e931d9e3..5890b5c1 100644
--- a/examples/rl/README.md
+++ b/examples/rl/README.md
@@ -56,6 +56,8 @@ cp -r megatron ../MindSpeed-MM/
 cd ..
 
 cd MindSpeed-MM
+# 安装mm-convert工具
+pip install -e . --no-deps
 mkdir -p logs data ckpt
 cd ..
 ```
@@ -97,8 +99,6 @@ cp -r mindspeed_rl ../MindSpeed-MM/
 cd ..
 
 cd MindSpeed-MM
-# 安装mm-convert工具
-pip install -e . --no-deps
 # MindSpeed/Megatron 中文件替换（必选）
 bash examples/rl/scripts/copy_adaptor_to_mindspeed.sh
 cd ..
-- 
Gitee


From 78553d7b00afd69c16207eef83470699b82ec155 Mon Sep 17 00:00:00 2001
From: mr-lin314 <798948055@qq.com>
Date: Sat, 30 Aug 2025 09:58:08 +0800
Subject: [PATCH 4/4] [docs]add mindspeed_rl commit id

---
 examples/rl/code/adaptor.py | 39 +++++++++++++++----------------------
 1 file changed, 16 insertions(+), 23 deletions(-)

diff --git a/examples/rl/code/adaptor.py b/examples/rl/code/adaptor.py
index 0c7660e4..5026fec1 100644
--- a/examples/rl/code/adaptor.py
+++ b/examples/rl/code/adaptor.py
@@ -29,6 +29,14 @@ def dot_product_attention_forward_impl(
         attention_bias: Tensor = None,
         packed_seq_params: Optional[PackedSeqParams] = None,
 ):
+    if packed_seq_params is None:
+        seq_length, bsz, n_head, head_dim = (
+            query.shape[0], query.shape[1], query.shape[2], query.shape[3]
+        )
+    else:
+        seq_length, n_head, head_dim = (
+            query.shape[0], query.shape[1], query.shape[2]
+        )
     use_remove_padding = getattr(self.config, 'use_remove_padding', False)
     if use_remove_padding:
         from mindspeed.utils import get_actual_seq_len
@@ -50,10 +58,6 @@ def dot_product_attention_forward_impl(
             sparse_mode=3)[0].reshape(seq_length, bsz, -1)
 
     else:
-        seq_length, bsz, n_head, head_dim = (
-            query.shape[0], query.shape[1], query.shape[2], query.shape[3]
-        )
-
         sparse_mode = self.config.sparse_mode
         if attn_mask_type == AttnMaskType.no_mask:
             sparse_mode = 0  # default mask
@@ -64,17 +68,15 @@ def dot_product_attention_forward_impl(
             else self.softmax_scale
         )
 
-        if packed_seq_params is not None:  # TND
-            actual_seq_qlen = packed_seq_params.cu_seqlens_q.tolist()
-            actual_seq_kvlen = packed_seq_params.cu_seqlens_kv.tolist()
-            query, key, value = (
-                [
-                    rearrange(x, 's b h d -> (b s) h d')
-                    for x in [query, key, value]
-                ]
-            )
+        if packed_seq_params is not None: # TND
+            if isinstance(packed_seq_params.cu_seqlens_q, list):
+                actual_seq_qlen = packed_seq_params.cu_seqlens_q
+                actual_seq_kvlen = packed_seq_params.cu_seqlens_kv
+            else:
+                actual_seq_qlen = packed_seq_params.cu_seqlens_q.tolist()
+                actual_seq_kvlen = packed_seq_params.cu_seqlens_kv.tolist()
             shape_order = 'TND'
-        else:  # SBH
+        else: # SBH
             actual_seq_qlen = None
             actual_seq_kvlen = None
             query, key, value = (
@@ -102,13 +104,4 @@ def dot_product_attention_forward_impl(
             actual_seq_kvlen=actual_seq_kvlen
         )[0]
 
-        if packed_seq_params is not None:
-            output = (
-                rearrange(
-                    output,
-                    '(b s) h d -> s b (h d)',
-                    s=seq_length, b=bsz
-                )
-            )
-
     return output
-- 
Gitee