[*.py] bash code_style.sh ; [MaxText/maxengine.py] Add pylint: disable=arguments-differ to MaxEngine::prefill_multisampling

Author: SamuelMarks

MaxText/checkpointing.py

@@ -103,7 +103,7 @@ def create_orbax_emergency_checkpoint_manager(
 
   # Only create directories if running on GPUs as the previous
   # directory structure might be assumed by TPUs
-  if global_mesh.devices.flatten()[0].platform == 'gpu':
+  if global_mesh.devices.flatten()[0].platform == "gpu":
     # pylint: disable=protected-access
     local_checkpoint_dir = f"{local_checkpoint_dir}/{jax._src.distributed.global_state.process_id}"
     local_p = epath.Path(local_checkpoint_dir)
@@ -374,9 +374,7 @@ def setup_checkpoint_logger(config) -> Any | None:  # pytype: disable=attribute-
   max_logging.log("Setting up checkpoint logger...")
   if config.enable_checkpoint_cloud_logger:
     logger_name = f"goodput_{config.run_name}"
-    options = ocp.logging.CloudLoggerOptions(
-        job_name=config.run_name, logger_name=logger_name
-    )
+    options = ocp.logging.CloudLoggerOptions(job_name=config.run_name, logger_name=logger_name)
     orbax_cloud_logger = ocp.logging.CloudLogger(options=options)
     max_logging.log("Successfully set up checkpoint cloud logger.")
     return orbax_cloud_logger

MaxText/decode.py

@@ -140,12 +140,12 @@ def main(argv: Sequence[str]) -> None:
   for i in range(_NUM_STREAMS):
     with jax.profiler.StepTraceAnnotation("prefill", stream=i):
       prefill_result, first_token = engine.prefill(
-        params=params,
-        padded_tokens=tokens,
-        images=processor_output.pixel_values,
-        true_length=true_length,
-        rng=rng_prefill,
-        slot=i,
+          params=params,
+          padded_tokens=tokens,
+          images=processor_output.pixel_values,
+          true_length=true_length,
+          rng=rng_prefill,
+          slot=i,
       )
     prefill_result_list.append(prefill_result)
     first_token_list.append(first_token)
@@ -178,6 +178,7 @@ def main(argv: Sequence[str]) -> None:
   # Deactivate profiler
   prof.deactivate()
 
+
 def _validate_config(config):
   assert config.load_full_state_path == "", (
       "Decode doesn't operate on full states! Convert to parameter checkpoint first." "Using generate_param_only_checkpoint."

MaxText/layers/linears.py

@@ -84,9 +84,7 @@ def __init__(
       axis: Union[Iterable[int], int] = -1,
       weight_dtype: DType = jnp.float32,
       dtype: DType = jnp.float32,
-      kernel_init: NdInitializer = nd_dense_init(
-          1.0, "fan_in", "truncated_normal"
-      ),
+      kernel_init: NdInitializer = nd_dense_init(1.0, "fan_in", "truncated_normal"),
       kernel_axes: Tuple[Optional[str], ...] = (),
       quant: Optional[Quant] = None,
       use_bias: bool = False,
@@ -127,9 +125,7 @@ def __init__(
     # Parameter initialization
     kernel_shape = self.in_features + self.out_features
     kernel_in_axis = np.arange(len(self.axis))
-    kernel_out_axis = np.arange(
-        len(self.axis), len(self.axis) + len(self.out_features)
-    )
+    kernel_out_axis = np.arange(len(self.axis), len(self.axis) + len(self.out_features))
 
     if not quantizations.in_serve_mode(self.quant):
       self.kernel = nnx.Param(
@@ -218,9 +214,7 @@ def dense_general(
     axis: Union[Iterable[int], int] = -1,
     weight_dtype: DType = jnp.float32,
     dtype: DType = jnp.float32,
-    kernel_init: NdInitializer = nd_dense_init(
-        1.0, "fan_in", "truncated_normal"
-    ),
+    kernel_init: NdInitializer = nd_dense_init(1.0, "fan_in", "truncated_normal"),
     kernel_axes: Tuple[Optional[str], ...] = (),
     quant: Optional[Quant] = None,
     use_bias: bool = False,
@@ -247,15 +241,11 @@ def dense_general(
     name: name passed to the ToLinen Module
   """
   if not (inputs_shape is not None) ^ (in_features is not None):
-    raise ValueError(
-        "Exactly one of inputs_shape or in_features must be specified."
-    )
+    raise ValueError("Exactly one of inputs_shape or in_features must be specified.")
 
   if inputs_shape is not None:
     axis = _canonicalize_tuple(axis)
-    in_features = tuple(
-        inputs_shape[ax] for ax in _normalize_axes(axis, len(inputs_shape))
-    )
+    in_features = tuple(inputs_shape[ax] for ax in _normalize_axes(axis, len(inputs_shape)))
   else:
     assert in_features is not None
   module = nnx.bridge.to_linen(
@@ -401,4 +391,3 @@ def __call__(self, inputs, decode: bool = False, deterministic: bool = False):
 
     output = checkpoint_name(output, "mlpwo")
     return output
-  

MaxText/layers/llama4.py

@@ -58,7 +58,7 @@ class Llama4UnfoldConvolution(nn.Module):
 
   def setup(self):
     """
-      Initialize Llama4UnfoldConvolution
+    Initialize Llama4UnfoldConvolution
     """
     cfg = self.config
     # Linear projection layer using dense_general.
@@ -190,7 +190,7 @@ class Llama4VisionMLP2(nn.Module):
 
   def setup(self):
     """
-      Initialize Llama4VisionMLP2
+    Initialize Llama4VisionMLP2
     """
     cfg = self.config
     self.fc1 = linears.dense_general(

MaxText/layers/models.py

@@ -666,9 +666,7 @@ def __call__(
           inputs_shape=y.shape,
           features=cfg.vocab_size,
           weight_dtype=cfg.weight_dtype,
-          dtype=jnp.float32
-          if cfg.logits_dot_in_fp32
-          else cfg.dtype,  # for logit training stability
+          dtype=jnp.float32 if cfg.logits_dot_in_fp32 else cfg.dtype,  # for logit training stability
           kernel_axes=("embed", "vocab"),
           name="logits_dense",
           matmul_precision=self.config.matmul_precision,
@@ -804,4 +802,3 @@ def __call__(
         image_embeddings=image_embeddings,
     )
     return logits
-  

MaxText/layers/multi_token_prediction.py

@@ -136,4 +136,3 @@ def __call__(
     # Shape: [B, S, H]
     # --- Return Processed Hidden State ---
     return next_hidden_state
-  

MaxText/maxengine.py

@@ -587,6 +587,7 @@ def prefill_multisampling_aot(  # pylint: disable=too-many-positional-arguments
 
   def prefill_multisampling(
       self,  # pytype: disable=signature-mismatch
+      # pylint: disable=arguments-differ
       *,
       params: Params,
       padded_tokens: jax.Array,

MaxText/tests/hf_checkpoint_conversion_check.py

@@ -13,6 +13,7 @@
 See the License for the specific language governing permissions and
 limitations under the License.
 """
+
 import os
 import torch
 import torch.nn.functional as F