Skip to content

Test: Improve model re-use in C++ DGX tests for CI stability #4263

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

Merged
merged 11 commits into from
May 19, 2025
Merged

Test: Improve model re-use in C++ DGX tests for CI stability #4263

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
68c2f79
Fix padded vocab size for Llama
DomBrown May 12, 2025
eda828d
Refactor multi GPU llama executor tests, and reuse the built model en…
DomBrown May 12, 2025
31d2830
Fix test list typo
DomBrown May 13, 2025
77465c6
WIP
DomBrown May 14, 2025
c619cc9
Further WIP
DomBrown May 15, 2025
a9db385
WIP
DomBrown May 15, 2025
f832956
Update test lists and readme
DomBrown May 15, 2025
b27e09c
Try parametrize for asymmetric
DomBrown May 16, 2025
3f6bd08
Parametrize + skip unsupported combinations
DomBrown May 16, 2025
6e7010c
Update test list
DomBrown May 16, 2025
e4998ee
Reduce environment duplicated code
DomBrown May 16, 2025
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
2 changes: 1 addition & 1 deletion cpp/tests/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,7 @@ pytest tests/integration/defs/cpp/test_e2e.py::test_model[llama-90]

pytest tests/integration/defs/cpp/test_e2e.py::test_benchmarks[gpt-90]

pytest tests/integration/defs/cpp/test_multi_gpu.py::test_disagg[90]
pytest tests/integration/defs/cpp/test_multi_gpu.py::TestDisagg::test_symmetric_executor[gpt-mpi_kvcache-90]
```

## Manual steps
Expand Down
20 changes: 12 additions & 8 deletions cpp/tests/executor/disaggExecutorTest.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,7 @@ namespace
{

auto constexpr LLAMA_INPUT_FILE = "input_tokens_llama.npy";
auto constexpr LLAMA_VOCAB_SIZE_PADDED = 128256;
auto constexpr LLAMA_END_ID = 128001;
auto constexpr LLAMA_PAD_ID = 128001;

Expand Down Expand Up @@ -520,6 +521,7 @@ TEST_P(DisaggParamsTest, DisaggTokenComparison)
// set defaults and adjust if needed by different models
fs::path inputPath = DATA_PATH / "input_tokens.npy";
ModelIds modelIds{50256, 50256};
SizeType32 vocabSizePadded{50257}; // gpt vocabSizePadded
bool isSpeculativeDecoding{false};

// NOTE: This can be used to disable checks for certain prompt batch entries
Expand Down Expand Up @@ -557,6 +559,7 @@ TEST_P(DisaggParamsTest, DisaggTokenComparison)
|| modelName == "llama_tp1_pp2_cp1" || modelName == "llama_tp2_pp1_cp1" || modelName == "llama_tp1_pp1_cp1")
{
inputPath = DATA_PATH / LLAMA_INPUT_FILE;
vocabSizePadded = LLAMA_VOCAB_SIZE_PADDED;

auto const resultsPath
= LLAMA_DATA_PATH / ((beamWidth == 1) ? "sampling" : "beam_search_" + std::to_string(beamWidth));
Expand Down Expand Up @@ -619,8 +622,6 @@ TEST_P(DisaggParamsTest, DisaggTokenComparison)
}
}

SizeType32 constexpr vocabSizePadded{50257}; // gpt vocabSizePadded

// Returning logits will bring higher latency
if (streaming && (outConfig.returnContextLogits || outConfig.returnGenerationLogits))
{
Expand Down Expand Up @@ -756,6 +757,7 @@ TEST_P(DisaggOrchestratorParamsTest, DisaggTokenComparison)
// set defaults and adjust if needed by different models
fs::path inputPath = DATA_PATH / "input_tokens.npy";
ModelIds modelIds{50256, 50256};
SizeType32 vocabSizePadded{50257}; // gpt vocabSizePadded
bool isSpeculativeDecoding{false};

// NOTE: This can be used to disable checks for certain prompt batch entries
Expand All @@ -764,6 +766,7 @@ TEST_P(DisaggOrchestratorParamsTest, DisaggTokenComparison)
|| modelName == "llama_tp1_pp2" || modelName == "llama_tp2_pp1" || modelName == "llama_tp1_pp1")
{
inputPath = DATA_PATH / LLAMA_INPUT_FILE;
vocabSizePadded = LLAMA_VOCAB_SIZE_PADDED;

auto const resultsPath
= LLAMA_DATA_PATH / ((beamWidth == 1) ? "sampling" : "beam_search_" + std::to_string(beamWidth));
Expand Down Expand Up @@ -827,8 +830,6 @@ TEST_P(DisaggOrchestratorParamsTest, DisaggTokenComparison)
}
}

SizeType32 constexpr vocabSizePadded{50257}; // gpt vocabSizePadded

// Returning logits will bring higher latency
if (streaming && (outConfig.returnContextLogits || outConfig.returnGenerationLogits))
{
Expand Down Expand Up @@ -920,6 +921,7 @@ TEST_P(ConditionalDisaggParamsTest, DisaggTokenComparison)
// set defaults and adjust if needed by different models
fs::path inputPath = DATA_PATH / "input_tokens.npy";
ModelIds modelIds{50256, 50256};
SizeType32 vocabSizePadded{50257}; // gpt vocabSizePadded
bool isSpeculativeDecoding{false};

// NOTE: This can be used to disable checks for certain prompt batch entries
Expand All @@ -935,6 +937,7 @@ TEST_P(ConditionalDisaggParamsTest, DisaggTokenComparison)
else if (modelName == "llama_tp1_pp1_cp1")
{
inputPath = DATA_PATH / LLAMA_INPUT_FILE;
vocabSizePadded = LLAMA_VOCAB_SIZE_PADDED;

auto const resultsPath
= LLAMA_DATA_PATH / ((beamWidth == 1) ? "sampling" : "beam_search_" + std::to_string(beamWidth));
Expand All @@ -948,8 +951,6 @@ TEST_P(ConditionalDisaggParamsTest, DisaggTokenComparison)
TLLM_THROW("Unrecognized modelName");
}

SizeType32 constexpr vocabSizePadded{50257}; // gpt vocabSizePadded

auto executorConfig = ExecutorConfig(maxBeamWidth);
FloatType freeGpuMemoryFraction = 0.9f;
KvCacheConfig kvCacheConfig{true, std::nullopt, std::nullopt, std::nullopt, freeGpuMemoryFraction};
Expand Down Expand Up @@ -1115,8 +1116,11 @@ INSTANTIATE_TEST_SUITE_P(GptSingleDeviceDisaggSymmetricExecutorMixedTest, Disagg
testing::Values(1)),
generateTestNameDisaggParams);

INSTANTIATE_TEST_SUITE_P(ConditionalDisaggSymmetricExecutorTest, ConditionalDisaggParamsTest,
testing::Combine(testing::Values("gpt", "llama_tp1_pp1_cp1")), generateTestNameCondDisaggParams);
INSTANTIATE_TEST_SUITE_P(GptConditionalDisaggSymmetricExecutorTest, ConditionalDisaggParamsTest,
testing::Combine(testing::Values("gpt")), generateTestNameCondDisaggParams);

INSTANTIATE_TEST_SUITE_P(LlamaConditionalDisaggSymmetricExecutorTest, ConditionalDisaggParamsTest,
testing::Combine(testing::Values("llama_tp1_pp1_cp1")), generateTestNameCondDisaggParams);

INSTANTIATE_TEST_SUITE_P(LlamaTP2DisaggSymmetricExecutorTest, DisaggParamsTest,
testing::Combine(testing::Values(4),
Expand Down
Loading