[feat]: oproj tensor parallelism in pure DP and graph-mode scenarios. (#2167)

### What this PR does / why we need it?
This PR introduces Oproj matrix tensor model parallel to achieve
decreasing of memory consumption. It only support graph mode in pure DP
scenario.

In deepseek r1 w8a8 PD disagregated Decode instance, using pure DP, with
oproj_tensor_parallel_size = 8, we have 1 ms TPOT increasing, saved 5.8
GB NPU memory per RANK. We got best performance when
oproj_tensor_parallel_size=4 without TPOT increasing.

performance data:
<img width="1442" height="442" alt="image"
src="https://github.com/user-attachments/assets/83270fc5-868a-4387-b0a9-fac29b4a376d"
/>

### Does this PR introduce _any_ user-facing change?
This PR introduces one new config in `additional_config`.
| Name | Effect | Required | Type | Constraints |
| :---------------------------- |
:--------------------------------------- | :------- | :--- |
:----------------- |
| oproj_tensor_parallel_size | Split the o_proj matrix along the row
dimension (head num * head dim) into oproj_tensor_parallel_size pieces.
| No | int | default value is None, once this value is set, the feature
will be enabled, head num * head dim must be divisible by this value. |

example

`--additional_config={"oproj_tensor_parallel_size": 8}`

### How was this patch tested?


- vLLM version: v0.10.1.1
- vLLM main:
eddaafc1c7

---------

Signed-off-by: zzhx1 <zzh_201018@outlook.com>
Co-authored-by: zzh <zzh_201018@outlook.com>

vllm_ascend/distributed/parallel_state.py

@@ -11,7 +11,7 @@ from vllm_ascend.ascend_config import get_ascend_config
 # Currently, mc2 op need their own group coordinator.
 _MC2: Optional[GroupCoordinator] = None
 _MLP_TP: Optional[GroupCoordinator] = None
-
+_OTP: Optional[GroupCoordinator] = None
 _LMTP: Optional[GroupCoordinator] = None
 
 
@@ -20,6 +20,12 @@ def get_mc2_group() -> GroupCoordinator:
     return _MC2
 
 
+def get_otp_group() -> GroupCoordinator:
+    assert _OTP is not None, (
+        "output tensor parallel group is not initialized")
+    return _OTP
+
+
 def get_lmhead_tp_group() -> GroupCoordinator:
     assert _LMTP is not None, (
         "lm head tensor parallel group is not initialized")
@@ -74,6 +80,20 @@ def init_ascend_model_parallel(parallel_config: ParallelConfig, ):
                                             backend,
                                             group_name="mlp_tp")
 
+    # If oproj tensor parallel size is set, we will create a group for it.
+    otp_size = get_ascend_config().oproj_tensor_parallel_size
+    if otp_size is not None:
+        group_ranks = []
+        global _OTP
+        num_oproj_tensor_parallel_groups: int = (world_size // otp_size)
+        for i in range(num_oproj_tensor_parallel_groups):
+            ranks = list(range(i * otp_size, (i + 1) * otp_size))
+            group_ranks.append(ranks)
+        _OTP = init_model_parallel_group(group_ranks,
+                                         get_world_group().local_rank,
+                                         backend,
+                                         group_name="otp")
+
     lmhead_tensor_parallel_size = get_ascend_config(
     ).lmhead_tensor_parallel_size
     if lmhead_tensor_parallel_size is not None:
@@ -117,3 +137,8 @@ def destroy_ascend_model_parallel():
     if _LMTP:
         _LMTP.destroy()
     _LMTP = None
+
+    global _OTP
+    if _OTP:
+        _OTP.destroy()
+    _OTP = None