format.sh

.jenkins/lm-eval-harness/test_lm_eval_correctness.py

@@ -76,18 +76,14 @@ def report_performance(task, input_lens, output_lens, time, record_property):
     context_lens = [i + o for i, o in zip(input_lens, output_lens)]
     gen_tput = sum(output_lens) / time
     all_lens = [input_lens, output_lens, context_lens]
-    min_input_tokens, min_output_tokens, min_context_tokens = [
-        min(x) for x in all_lens
-    ]
-    max_input_tokens, max_output_tokens, max_context_tokens = [
-        max(x) for x in all_lens
-    ]
-    mean_input_tokens, mean_output_tokens, mean_context_tokens = [
-        statistics.mean(x) for x in all_lens
-    ]
-    stddev_input_tokens, stddev_output_tokens, stddev_context_tokens = [
-        statistics.stdev(x) for x in all_lens
-    ]
+    min_input_tokens, min_output_tokens, min_context_tokens = (
+        min(x) for x in all_lens)
+    max_input_tokens, max_output_tokens, max_context_tokens = (
+        max(x) for x in all_lens)
+    mean_input_tokens, mean_output_tokens, mean_context_tokens = (
+        statistics.mean(x) for x in all_lens)
+    stddev_input_tokens, stddev_output_tokens, stddev_context_tokens = (
+        statistics.stdev(x) for x in all_lens)
     msg = (
         f'{task} | estimated average generation throughput: {gen_tput:.2f} tokens/s \n'  # noqa: G004, E501
         f'{task} | input_tokens   | min: {min_input_tokens} | max: {max_input_tokens} | mean: {mean_input_tokens:.2f} | stddev: {stddev_input_tokens:.2f}\n'  # noqa: E501

setup.py

@@ -382,8 +382,7 @@ def get_gaudi_sw_version():
     output = subprocess.run("hl-smi",
                             shell=True,
                             text=True,
-                            stdout=subprocess.PIPE,
-                            stderr=subprocess.PIPE,
+                            capture_output=True,
                             env={"ENABLE_CONSOLE": "true"})
     if output.returncode == 0 and output.stdout:
         return output.stdout.split("\n")[2].replace(

vllm/executor/ray_hpu_executor.py

@@ -34,7 +34,7 @@ class RayHPUExecutor(DistributedGPUExecutor):
     uses_ray: bool = True
 
     def _init_executor(self) -> None:
-        self.forward_dag: Optional["ray.dag.CompiledDAG"] = None
+        self.forward_dag: Optional[ray.dag.CompiledDAG] = None
         # If the env var is set, it uses the Ray's compiled DAG API
         # which optimizes the control plane overhead.
         # Run vLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.

vllm/model_executor/layers/sampler.py

@@ -358,7 +358,7 @@ def _get_bin_counts_and_mask(
     return bin_counts, mask
 
 
-class ApplyToppTopkScalar():
+class ApplyToppTopkScalar:
     """
     The original implementation of _apply_top_k_top_p is more general
     as it uses vector topp, topk

vllm/worker/hpu_model_runner.py

@@ -73,8 +73,7 @@ class Singleton(type):
 
     def __call__(cls, *args, **kwargs):
         if cls not in cls._instances:
-            cls._instances[cls] = super(Singleton,
-                                        cls).__call__(*args, **kwargs)
+            cls._instances[cls] = super().__call__(*args, **kwargs)
         return cls._instances[cls]
 
 
@@ -281,7 +280,7 @@ def precompute_indices_and_offsets(block_size, slot_mapping, is_prompt):
     return indices, offsets
 
 
-class HpuModelAdapter():
+class HpuModelAdapter:
 
     def __init__(self, model, block_size, dtype, enforce_eager):
         self.model = model
@@ -1788,7 +1787,7 @@ def _maybe_wrap_in_hpu_graph(*args, **kwargs):
     ) if htorch.utils.internal.is_lazy() else HpuModelAdapter(*args, **kwargs)
 
 
-class HabanaProfilerCounterHelper():
+class HabanaProfilerCounterHelper:
 
     def __init__(self):
         self.niter = 0