Merge branch 'habana_main' into dev/mfylcek/sampler-aware_batch_size_…

…padding
HabanaAI · Jan 10, 2025 · 2520ec5 · 2520ec5
2 parents deda42a + 73aaf71
commit 2520ec5
Show file tree

Hide file tree

Showing 17 changed files with 459 additions and 238 deletions.
diff --git a/.jenkins/test_config.yaml b/.jenkins/test_config.yaml
@@ -57,4 +57,4 @@ stages:
         command: TORCH_COMPILE_DISABLE=true VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True pytest -v tests/spec_decode/e2e/test_medusa_correctness.py::test_medusa_e2e_greedy_correctness
       - name: gsm8k_small_g2_tp1_eagle_spec_decode
         flavor: g2
-        command: TORCH_COMPILE_DISABLE=true VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True pytest -v tests/spec_decode/e2e/test_eagle_correctness.py::test_eagle_e2e_greedy_correctness
+        command: VLLM_COS_SIN_RECOMPUTE=true TORCH_COMPILE_DISABLE=true VLLM_CONTIGUOUS_PA=false VLLM_SKIP_WARMUP=True pytest -v tests/spec_decode/e2e/test_eagle_correctness.py::test_eagle_e2e_greedy_correctness
diff --git a/benchmarks/benchmark_throughput.py b/benchmarks/benchmark_throughput.py
@@ -57,6 +57,10 @@ def _get_prompt_for_image_model(question: str, *, model: str) -> str:
     model = model.lower()
     if "pixtral" in model:
         return f"<s>[INST]{question}\n[IMG][/INST]"
+    elif "llava" in model:
+        return f"USER: <image>\n{question}\nASSISTANT:"
+    elif "llama-3.2" in model:
+        return f"<|image|><|begin_of_text|>{question}"
     raise ValueError(f"Unsupported model {model}")
 
 

diff --git a/tests/lora/conftest.py b/tests/lora/conftest.py
@@ -64,6 +64,7 @@ def cleanup_fixture(should_do_global_cleanup_after_test: bool):
 
 @pytest.fixture
 def dist_init():
+    import habana_frameworks.torch.hpu  # noqa: F401
     temp_file = tempfile.mkstemp()[1]
     backend_type = "hccl" if current_platform.is_hpu() else "nccl"
     init_distributed_environment(