EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[Disaggregated Prefill] P2P Disaggregated Prefill based on llm_datadist (#694)

Browse Source 
### What this PR does / why we need it?
- This PR proposes a P2P version of Disaggregated Prefill based on
llm_datadist which manages data transfer.

- This solution reconstructs previous offline single-node Disaggregated
Prefill solution, and supports multi-node and online serveing now.

- Currently this solution supports 1P1D situation of Deepseek hybrid
parallelism (P: TP+EP, D: DP+EP). Note that xPyD situation is considered
in the solution design, and will be supported soon within v1 engine.

---------

Signed-off-by: hw_whx <wanghexiang7@huawei.com>
Signed-off-by: ganyi <pleaplusone.gy@gmail.com>
Co-authored-by: hw_whx <wanghexiang7@huawei.com>
Co-authored-by: ganyi <pleaplusone.gy@gmail.com>
This commit is contained in:
whx
2025-05-01 22:31:36 +08:00
committed by GitHub
parent 84e2ed898b
commit 8b194ad12e
18 changed files with 1769 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
tests/singlecard/spec_decode/e2e/test_mtp_correctness.py
View File

@@ -123,7 +123,7 @@ def test_mtp_e2e_greedy_correctness(vllm_runner, common_llm_kwargs,
"model_name": QUANT_MODEL,
# GPU memory utilization
"gpu_memory_utilization": 0.85
"gpu_memory_utilization": 0.8
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -169,7 +169,7 @@ def test_mtp_e2e_quant_greedy_correctness(vllm_runner, common_llm_kwargs,
"model_name": FLOAT_MODEL,
# GPU memory utilization
"gpu_memory_utilization": 0.85
"gpu_memory_utilization": 0.8
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -230,7 +230,7 @@ def test_mtp_e2e_greedy_logprobs(vllm_runner, common_llm_kwargs,
# Main model
"model_name": FLOAT_MODEL,
"gpu_memory_utilization": 0.85
"gpu_memory_utilization": 0.8
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -274,7 +274,7 @@ def test_mtp_e2e_greedy_correctness_torchair_graph(
# Main model
"model_name": QUANT_MODEL,
"gpu_memory_utilization": 0.85
"gpu_memory_utilization": 0.8
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -322,7 +322,7 @@ def test_mtp_e2e_quant_greedy_correctness_torchair_graph(
"model_name": FLOAT_MODEL,
# GPU memory utilization
"gpu_memory_utilization": 0.9
"gpu_memory_utilization": 0.8
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -369,7 +369,7 @@ def test_mtp_e2e_greedy_correctness_with_preemption(
"model_name": FLOAT_MODEL,
# GPU memory utilization
"gpu_memory_utilization": 0.9
"gpu_memory_utilization": 0.8
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])
@@ -420,7 +420,7 @@ def test_mtp_different_k(vllm_runner, common_llm_kwargs,
"model_name": FLOAT_MODEL,
# GPU memory utilization
"gpu_memory_utilization": 0.9
"gpu_memory_utilization": 0.8
}])
@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
@pytest.mark.parametrize("baseline_llm_kwargs", [{}])