Skip to content

[TRTLLM-9523][chore] e2e tests for KV manager v2 with dis-agg #12689

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
Shixiaowei02 wants to merge 2 commits into from
+153 −1
Closed

[TRTLLM-9523][chore] e2e tests for KV manager v2 with dis-agg #12689

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
85a74db
add tests for KV manager with dis-agg
Shixiaowei02 Apr 2, 2026
7a4ade9
update
Shixiaowei02 Apr 2, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 12 additions & 1 deletion tensorrt_llm/_torch/pyexecutor/scheduler/scheduler_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -230,8 +230,19 @@ def _schedule_loop(self, active_requests, inflight_request_ids):

req_state_value = req.state_value

# Disagg gen init bypasses normal state gating (same as C++ / V1 scheduler)
# Disagg gen init bypasses normal state gating (same as C++ / V1 scheduler),
# but the V2 scheduler owns inline KV allocation so we must allocate here.
# V1 defers allocation to prepare_resources; V2 prepare_resources is a no-op
# for the primary manager, so allocation must happen in the scheduling loop.
if req_state_value == self._disagg_gen_init_state_value:
if not self.kv_cache_manager.prepare_context(req):
req_it += 1
continue
if not self.kv_cache_manager.resize_context(
req, req.context_remaining_length + get_draft_token_length(req)
):
req_it += 1
continue
disagg_candidates.append(req)
req_it += 1
continue
Expand Down
135 changes: 135 additions & 0 deletions tests/integration/defs/accuracy/test_disaggregated_serving.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -637,6 +637,50 @@ def test_auto_dtype(self, ctx_disable_overlap_scheduler,
self.MODEL_PATH) as llm:
run_accuracy_test(llm, self.MODEL_NAME, ["MMLU", "GSM8K"])

@skip_pre_hopper
@pytest.mark.skip_less_device(2)
def test_kv_cache_v2_nixl_python(self):
"""Test with use_kv_cache_manager_v2=True, block_reuse=False, backend=NIXL, transceiver_runtime=PYTHON"""
ctx_server_config = {
"disable_overlap_scheduler": True,
"kv_cache_config": {
"enable_block_reuse": False,
"use_kv_cache_manager_v2": True
},
"cache_transceiver_config": {
"backend": "NIXL",
"transceiver_runtime": "PYTHON"
}
}
gen_server_config = {
"disable_overlap_scheduler": False,
"kv_cache_config": {
"enable_block_reuse": False,
"use_kv_cache_manager_v2": True
},
"cache_transceiver_config": {
"backend": "NIXL",
"transceiver_runtime": "PYTHON"
}
}
disaggregated_server_config = {
"hostname": "localhost",
"port": 8000,
"backend": "pytorch",
"context_servers": {
"num_instances": 1,
"urls": ["localhost:8001"]
},
"generation_servers": {
"num_instances": 1,
"urls": ["localhost:8002"]
}
}
with launch_disaggregated_llm(disaggregated_server_config,
ctx_server_config, gen_server_config,
self.MODEL_PATH) as llm:
run_accuracy_test(llm, self.MODEL_NAME, ["GSM8K"])

@pytest.mark.skip_less_device(2)
def test_ngram(self):
speculative_decoding_config = {
Expand Down Expand Up @@ -1145,6 +1189,51 @@ def test_guided_decoding(self, backend: str, mtp_nextn: int, mocker):
self.MODEL_PATH) as llm:
run_accuracy_test(llm, self.MODEL_NAME, ["JsonModeEval"])

@pytest.mark.skip_less_device(2)
@pytest.mark.skip_less_device_memory(60000)
@skip_pre_hopper
def test_kv_cache_v2_nixl_python(self):
"""Test with use_kv_cache_manager_v2=True, block_reuse=False, backend=NIXL, transceiver_runtime=PYTHON"""
ctx_server_config = {
"disable_overlap_scheduler": True,
"kv_cache_config": {
"enable_block_reuse": False,
"use_kv_cache_manager_v2": True
},
"cache_transceiver_config": {
"backend": "NIXL",
"transceiver_runtime": "PYTHON"
}
}
gen_server_config = {
"disable_overlap_scheduler": True,
"kv_cache_config": {
"enable_block_reuse": False,
"use_kv_cache_manager_v2": True
},
"cache_transceiver_config": {
"backend": "NIXL",
"transceiver_runtime": "PYTHON"
}
}
disaggregated_server_config = {
"hostname": "localhost",
"port": 8000,
"backend": "pytorch",
"context_servers": {
"num_instances": 1,
"urls": ["localhost:8001"]
},
"generation_servers": {
"num_instances": 1,
"urls": ["localhost:8002"]
}
}
with launch_disaggregated_llm(disaggregated_server_config,
ctx_server_config, gen_server_config,
self.MODEL_PATH) as llm:
run_accuracy_test(llm, self.MODEL_NAME, ["GSM8K"])


@pytest.mark.timeout(DEFAULT_TEST_TIMEOUT)
class TestGemma3_1BInstruct(LlmapiAccuracyTestHarness):
Expand Down Expand Up @@ -1197,6 +1286,52 @@ def test_auto_dtype(self, block_reuse):
self.MODEL_PATH) as llm:
run_accuracy_test(llm, self.MODEL_NAME, ["MMLU", "GSM8K"])

@pytest.mark.skip_less_device(2)
@skip_pre_hopper
def test_kv_cache_v2_nixl_python(self):
"""Test with use_kv_cache_manager_v2=True, block_reuse=False, backend=NIXL, transceiver_runtime=PYTHON"""
ctx_server_config = {
"disable_overlap_scheduler": True,
"cuda_graph_config": None,
"kv_cache_config": {
"enable_block_reuse": False,
"use_kv_cache_manager_v2": True
},
"cache_transceiver_config": {
"backend": "NIXL",
"transceiver_runtime": "PYTHON"
}
}
gen_server_config = {
"disable_overlap_scheduler": True,
"cuda_graph_config": None,
"kv_cache_config": {
"enable_block_reuse": False,
"use_kv_cache_manager_v2": True
},
"cache_transceiver_config": {
"backend": "NIXL",
"transceiver_runtime": "PYTHON"
}
}
disaggregated_server_config = {
"hostname": "localhost",
"port": 8000,
"backend": "pytorch",
"context_servers": {
"num_instances": 1,
"urls": ["localhost:8001"]
},
"generation_servers": {
"num_instances": 1,
"urls": ["localhost:8002"]
}
}
with launch_disaggregated_llm(disaggregated_server_config,
ctx_server_config, gen_server_config,
self.MODEL_PATH) as llm:
run_accuracy_test(llm, self.MODEL_NAME, ["MMLU", "GSM8K"])


@skip_pre_blackwell
@pytest.mark.skip_less_device_memory(80000)
Expand Down
3 changes: 3 additions & 0 deletions tests/integration/test_lists/test-db/l0_dgx_b300.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,6 +92,9 @@ l0_dgx_b300:
- disaggregated/test_disaggregated.py::test_disaggregated_benchmark_on_diff_backends[DeepSeek-V3-Lite-fp8]
- accuracy/test_disaggregated_serving.py::TestQwen3_8B::test_nixl_backend
- accuracy/test_disaggregated_serving.py::TestDeepSeekV3Lite::test_nixl_backend
- accuracy/test_disaggregated_serving.py::TestDeepSeekV3Lite::test_kv_cache_v2_nixl_python
Comment thread
Shixiaowei02 marked this conversation as resolved.
- accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_kv_cache_v2_nixl_python
- accuracy/test_disaggregated_serving.py::TestGemma3_1BInstruct::test_kv_cache_v2_nixl_python
- accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_tp4] TIMEOUT (180)
- accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_pp4_mtp] TIMEOUT (180)
- condition:
Expand Down
3 changes: 3 additions & 0 deletions tests/integration/test_lists/test-db/l0_dgx_h100.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,9 +31,12 @@ l0_dgx_h100:
- accuracy/test_disaggregated_serving.py::TestQwen3_8B::test_chunked_prefill
- accuracy/test_disaggregated_serving.py::TestQwen3_8B::test_nixl_backend
- accuracy/test_disaggregated_serving.py::TestDeepSeekV3Lite::test_nixl_backend
- accuracy/test_disaggregated_serving.py::TestDeepSeekV3Lite::test_kv_cache_v2_nixl_python
- accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_ngram
- accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_kv_cache_v2_nixl_python
- accuracy/test_disaggregated_serving.py::TestGemma3_1BInstruct::test_auto_dtype[False]
- accuracy/test_disaggregated_serving.py::TestGemma3_1BInstruct::test_auto_dtype[True]
- accuracy/test_disaggregated_serving.py::TestGemma3_1BInstruct::test_kv_cache_v2_nixl_python
- accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_auto_dtype[False-False-False-False]
- accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_auto_dtype[False-False-False-True]
- accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_auto_dtype[False-False-True-False]
Expand Down
Loading