biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Compile] Conditional compilation. Introduce compile_ranges (#24252)

Browse Source 
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: Luka Govedič <luka.govedic@gmail.com>
Signed-off-by: ProExpertProg <lgovedic@redhat.com>
Co-authored-by: Luka Govedič <lgovedic@redhat.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Luka Govedič <luka.govedic@gmail.com>
This commit is contained in:
Ilya Markov
2025-12-05 19:17:32 +01:00
committed by GitHub
parent 66e674cdd5
commit 4e26d3b09e
15 changed files with 582 additions and 268 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
tests/compile/distributed/test_fusions_e2e.py
View File

@@ -298,10 +298,14 @@ def test_tp2_attn_quant_allreduce_rmsnorm(
r"fusion_attn.py:\d+] Fused quant onto (\d+) attention nodes",
log_holder.text,
)
assert len(log_matches) == 2, log_holder.text
# 2 for each compile range
# (global compile range can be split due to fuse_allreduce_rmsnorm)
num_compile_ranges = len(compilation_config.get_compile_ranges())
assert num_compile_ranges in [1, 2]
assert int(log_matches[0]) == matches.attention_fusion
assert int(log_matches[1]) == matches.attention_fusion
assert len(log_matches) == 2 * num_compile_ranges, log_holder.text
assert all(int(log_match) == matches.attention_fusion for log_match in log_matches)
log_matches = re.findall(
r"collective_fusion.py:\d+] Replaced (\d+) patterns",
@@ -312,6 +316,12 @@ def test_tp2_attn_quant_allreduce_rmsnorm(
assert int(log_matches[0]) == matches.allreduce_fusion
assert int(log_matches[1]) == matches.allreduce_fusion
log_matches = re.findall(
r"pass_manager.py:\d+] Skipping .*AllReduceFusionPass.* with compile range",
log_holder.text,
)
assert len(log_matches) == 2 * (num_compile_ranges - 1), log_holder.text
@multi_gpu_test(num_gpus=2)
@pytest.mark.parametrize(
@@ -446,7 +456,6 @@ def run_model(compile_config: int | CompilationConfig, model: str, **model_kwarg
# No cudagraphs by default
if compilation_config.cudagraph_mode is None:
compilation_config.cudagraph_mode = CUDAGraphMode.NONE
llm = LLM(
model=model,
compilation_config=compilation_config,
@@ -459,3 +468,9 @@ def run_model(compile_config: int | CompilationConfig, model: str, **model_kwarg
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
# Get the compile ranges split points after vllm config post init
# in order to compute compile ranges correctly
compilation_config.compile_ranges_split_points = (
llm.llm_engine.vllm_config.compilation_config.compile_ranges_split_points
)