[npu] Combine NPU test fixes from #35472 and #34516 (#36180)

test/registered/npu/accuracy/qwen3_vl_30b_a3b_thinking/test_npu_qwen3_vl_30b_a3b_thinking_1p_mmmu.py

@@ -20,7 +20,7 @@ ENVS = {
     "HCCL_SOCKET_IFNAME": "lo",
     "GLOO_SOCKET_IFNAME": "lo",
     "HCCL_OP_EXPANSION_MODE": "AIV",
-    "HCCL_BUFFSIZE": "2000",
+    "DEEPEP_HCCL_BUFFSIZE": "2000",
 }
 
 OTHER_ARGS = [
@@ -54,7 +54,7 @@ OTHER_ARGS = [
 ]
 
 
-class TestQwen3(TestNpuAccuracyTestCaseBase):
+class TestQwen3_VL_30B_A3B_Thinking_MMMU(TestNpuAccuracyTestCaseBase):
     model = QWEN3_VL_30B_A3B_THINKING_MODEL_PATH
     envs = ENVS
     other_args = OTHER_ARGS

test/registered/npu/accuracy/qwen3_vl_8b_thinking/test_npu_qwen3_vl_8b_thinking_1p_mmmu.py

@@ -10,7 +10,7 @@ from sglang.test.ascend.e2e.test_npu_performance_utils import (
 from sglang.test.ci.ci_register import register_npu_ci
 
 register_npu_ci(est_time=3600, suite="base-c-test-perf-2-npu-a3")
-register_npu_ci(est_time=6500, suite="nightly-acc-2-npu-a3", nightly=True)
+register_npu_ci(est_time=12000, suite="nightly-acc-2-npu-a3", nightly=True)
 
 _is_pr_pipeline = os.environ.get("GITHUB_EVENT_NAME") == "pull_request"
 
@@ -20,7 +20,7 @@ ENVS = {
     "HCCL_SOCKET_IFNAME": "lo",
     "GLOO_SOCKET_IFNAME": "lo",
     "HCCL_OP_EXPANSION_MODE": "AIV",
-    "HCCL_BUFFSIZE": "2000",
+    "DEEPEP_HCCL_BUFFSIZE": "2000",
 }
 
 OTHER_ARGS = [
@@ -52,7 +52,7 @@ OTHER_ARGS = [
 ]
 
 
-class TestQwen3(TestNpuAccuracyTestCaseBase):
+class TestQwen3_VL_8B_Thinking_MMMU(TestNpuAccuracyTestCaseBase):
     model = QWEN3_VL_8B_THINKING_MODEL_PATH
     envs = ENVS
     other_args = OTHER_ARGS

test/registered/npu/performance/kimi_k2_6/test_npu_kimi_k2_6_w4a8_8p_in3k5_out1k5_20ms.py

@@ -53,6 +53,8 @@ KIMI_K2_6_OTHER_ARGS = [
     6144,
     "--max-prefill-tokens",
     65536,
+    "--max-total-tokens",
+    32256,
     "--enable-multimodal",
     "--mm-attention-backend",
     "ascend_attn",