[None][perf] Validate GPT-OSS transceiver v2 performance

tests/integration/defs/accuracy/test_disaggregated_serving.py

@@ -1568,15 +1568,17 @@ def test_auto_dtype(self, block_reuse, mocker):
         ctx_server_config = {
             "disable_overlap_scheduler": True,
             "cache_transceiver_config": {
-                "backend": "DEFAULT",
+                "backend": "NIXL",
+                "transceiver_runtime": "PYTHON",
                 "max_tokens_in_buffer": 4096
             },
             "tensor_parallel_size": 4
         }
         gen_server_config = {
             "disable_overlap_scheduler": False,
             "cache_transceiver_config": {
-                "backend": "DEFAULT",
+                "backend": "NIXL",
+                "transceiver_runtime": "PYTHON",
                 "max_tokens_in_buffer": 4096
             },
             "tensor_parallel_size": 4

tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_gptoss_eagle_triton.yaml

@@ -26,7 +26,8 @@ context_servers:
   cuda_graph_config: null
   print_iter_log: true
   cache_transceiver_config:
-    backend: DEFAULT
+    backend: NIXL
+    transceiver_runtime: PYTHON
     max_tokens_in_buffer: 16384
 generation_servers:
   num_instances: 1
@@ -63,5 +64,6 @@ generation_servers:
     - 1024
   print_iter_log: true
   cache_transceiver_config:
-    backend: DEFAULT
+    backend: NIXL
+    transceiver_runtime: PYTHON
     max_tokens_in_buffer: 16384

tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_gptoss_eagle_trtllm.yaml

@@ -26,7 +26,8 @@ context_servers:
   cuda_graph_config: null
   print_iter_log: true
   cache_transceiver_config:
-    backend: DEFAULT
+    backend: NIXL
+    transceiver_runtime: PYTHON
     max_tokens_in_buffer: 16384
 generation_servers:
   num_instances: 1
@@ -63,5 +64,6 @@ generation_servers:
     - 1024
   print_iter_log: true
   cache_transceiver_config:
-    backend: DEFAULT
+    backend: NIXL
+    transceiver_runtime: PYTHON
     max_tokens_in_buffer: 16384

tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_gptoss_tllm.yaml

@@ -22,7 +22,8 @@ context_servers:
   cuda_graph_config: null
   print_iter_log: true
   cache_transceiver_config:
-    backend: DEFAULT
+    backend: NIXL
+    transceiver_runtime: PYTHON
     max_tokens_in_buffer: 16384
 generation_servers:
   num_instances: 1
@@ -59,5 +60,6 @@ generation_servers:
     - 1024
   print_iter_log: true
   cache_transceiver_config:
-    backend: DEFAULT
+    backend: NIXL
+    transceiver_runtime: PYTHON
     max_tokens_in_buffer: 16384

tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_gptoss_triton.yaml

@@ -21,7 +21,8 @@ context_servers:
   cuda_graph_config: null
   print_iter_log: true
   cache_transceiver_config:
-    backend: DEFAULT
+    backend: NIXL
+    transceiver_runtime: PYTHON
     max_tokens_in_buffer: 16384
 generation_servers:
   num_instances: 1
@@ -53,5 +54,6 @@ generation_servers:
     - 128
   print_iter_log: true
   cache_transceiver_config:
-    backend: DEFAULT
+    backend: NIXL
+    transceiver_runtime: PYTHON
     max_tokens_in_buffer: 16384

tests/scripts/perf-sanity/disaggregated/gb200_gpt-oss-120b-fp4_1k1k_con2048_ctx1_tp1_gen1_dep2_eplb0_mtp0_ccb-NIXL.yaml

@@ -57,6 +57,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 1536
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -65,6 +66,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     num_postprocess_workers: 4
     stream_interval: 20
@@ -79,6 +81,7 @@ worker_config:
     enable_attention_dp: false
     cuda_graph_config: null
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -88,4 +91,5 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true

tests/scripts/perf-sanity/disaggregated/gb200_gpt-oss-120b-fp4_1k1k_con512_ctx1_tp1_gen1_dep2_eplb0_mtp0_ccb-NIXL.yaml

@@ -57,6 +57,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 1536
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -65,6 +66,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     num_postprocess_workers: 4
     stream_interval: 20
@@ -79,6 +81,7 @@ worker_config:
     enable_attention_dp: false
     cuda_graph_config: null
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -88,4 +91,5 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true

tests/scripts/perf-sanity/disaggregated/gb200_gpt-oss-120b-fp4_1k1k_con64_ctx1_tp1_gen1_tp4_eplb0_mtp0_ccb-NIXL.yaml

@@ -53,6 +53,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 256
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -61,6 +62,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     num_postprocess_workers: 4
     stream_interval: 20
@@ -75,6 +77,7 @@ worker_config:
     enable_attention_dp: false
     cuda_graph_config: null
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -84,4 +87,5 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true

tests/scripts/perf-sanity/disaggregated/gb200_gpt-oss-120b-fp4_8k1k_con1024_ctx1_tp1_gen1_tp4_eplb0_mtp0_ccb-NIXL.yaml

@@ -65,6 +65,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 1280
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.85
       dtype: fp8
@@ -73,6 +74,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 9216
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: false
     num_postprocess_workers: 4
     stream_interval: 20
@@ -91,6 +93,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 30
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.8
       dtype: fp8
@@ -99,6 +102,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 9216
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     num_postprocess_workers: 4
     stream_interval: 20

tests/scripts/perf-sanity/disaggregated/gb200_gpt-oss-120b-fp4_8k1k_con128_ctx1_tp1_gen1_tp4_eplb0_mtp0_ccb-NIXL.yaml

@@ -53,6 +53,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 1024
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -61,6 +62,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     num_postprocess_workers: 4
     stream_interval: 20
@@ -75,6 +77,7 @@ worker_config:
     enable_attention_dp: false
     cuda_graph_config: null
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -84,4 +87,5 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true

tests/scripts/perf-sanity/disaggregated/gb200_gpt-oss-120b-fp4_8k1k_con4_ctx1_tp1_gen1_tp4_eplb0_mtp0_ccb-NIXL.yaml

@@ -53,6 +53,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 1024
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -61,6 +62,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     num_postprocess_workers: 4
     stream_interval: 20
@@ -75,6 +77,7 @@ worker_config:
     enable_attention_dp: false
     cuda_graph_config: null
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -84,4 +87,5 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true

tests/scripts/perf-sanity/disaggregated/gb200_gpt-oss-120b-fp4_8k1k_con512_ctx1_tp1_gen1_dep2_eplb0_mtp0_ccb-NIXL.yaml

@@ -57,6 +57,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 512
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -65,6 +66,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     num_postprocess_workers: 4
     stream_interval: 20
@@ -79,6 +81,7 @@ worker_config:
     enable_attention_dp: false
     cuda_graph_config: null
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -88,4 +91,5 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 8448
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true

tests/scripts/perf/disaggregated/gb200_gpt-oss-120b-fp4_1k1k_con2048_ctx1_tp1_gen1_dep2_eplb0_mtp0_ccb-NIXL.yaml

@@ -68,6 +68,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 1536
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -76,6 +77,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     num_postprocess_workers: 4
     stream_interval: 20
@@ -90,6 +92,7 @@ worker_config:
     enable_attention_dp: false
     cuda_graph_config: null
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -99,4 +102,5 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true

tests/scripts/perf/disaggregated/gb200_gpt-oss-120b-fp4_1k1k_con512_ctx1_tp1_gen1_dep2_eplb0_mtp0_ccb-NIXL.yaml

@@ -68,6 +68,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 1536
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -76,6 +77,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     num_postprocess_workers: 4
     stream_interval: 20
@@ -90,6 +92,7 @@ worker_config:
     enable_attention_dp: false
     cuda_graph_config: null
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -99,4 +102,5 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true

tests/scripts/perf/disaggregated/gb200_gpt-oss-120b-fp4_1k1k_con64_ctx1_tp1_gen1_tp4_eplb0_mtp0_ccb-NIXL.yaml

@@ -64,6 +64,7 @@ worker_config:
       enable_padding: true
       max_batch_size: 256
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -72,6 +73,7 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true
     num_postprocess_workers: 4
     stream_interval: 20
@@ -86,6 +88,7 @@ worker_config:
     enable_attention_dp: false
     cuda_graph_config: null
     kv_cache_config:
+      use_kv_cache_manager_v2: false
       enable_block_reuse: false
       free_gpu_memory_fraction: 0.9
       dtype: fp8
@@ -95,4 +98,5 @@ worker_config:
     cache_transceiver_config:
       max_tokens_in_buffer: 1024
       backend: NIXL
+      transceiver_runtime: PYTHON
     disable_overlap_scheduler: true