Decouple Filter MP Rules function from cuda imports

plugins/fused-ops-and-kernels/src/fms_acceleration_foak/framework_plugin_fast_kernels.py

@@ -23,8 +23,7 @@
 import torch
 
 # Local
-from .models.utils import filter_mp_rules
-from .utils import lora_adapters_switch_ddp_from_fsdp
+from .utils import filter_mp_rules, lora_adapters_switch_ddp_from_fsdp
 
 
 # consider rewriting register_foak_model_patch_rules into something

plugins/fused-ops-and-kernels/src/fms_acceleration_foak/models/granite.py

@@ -30,12 +30,12 @@
 from ..kernels.unsloth.cross_entropy_loss import FastCrossEntropyLoss
 from ..kernels.unsloth.rms_layernorm import fast_rms_layernorm
 from ..kernels.unsloth.rope_embedding import fast_rope_embedding
+from ..utils import filter_mp_rules
 from .utils import (
     KEY_MLP,
     KEY_O,
     KEY_QKV,
     build_lora_fused_ops,
-    filter_mp_rules,
     get_hidden_activation_fn_key,
     trigger_fused_ops,
 )

plugins/fused-ops-and-kernels/src/fms_acceleration_foak/models/llama.py

@@ -36,12 +36,12 @@
 from ..kernels.unsloth.cross_entropy_loss import FastCrossEntropyLoss
 from ..kernels.unsloth.rms_layernorm import fast_rms_layernorm
 from ..kernels.unsloth.rope_embedding import fast_rope_embedding
+from ..utils import filter_mp_rules
 from .utils import (
     KEY_MLP,
     KEY_O,
     KEY_QKV,
     build_lora_fused_ops,
-    filter_mp_rules,
     get_hidden_activation_fn_key,
     trigger_fused_ops,
 )

plugins/fused-ops-and-kernels/src/fms_acceleration_foak/models/mistral.py

@@ -36,12 +36,12 @@
 from ..kernels.unsloth.cross_entropy_loss import FastCrossEntropyLoss
 from ..kernels.unsloth.rms_layernorm import fast_rms_layernorm
 from ..kernels.unsloth.rope_embedding import fast_rope_embedding
+from ..utils import filter_mp_rules
 from .utils import (
     KEY_MLP,
     KEY_O,
     KEY_QKV,
     build_lora_fused_ops,
-    filter_mp_rules,
     get_hidden_activation_fn_key,
     trigger_fused_ops,
 )

plugins/fused-ops-and-kernels/src/fms_acceleration_foak/models/utils.py

@@ -1,10 +1,10 @@
 # Standard
 from functools import partial
-from typing import Callable, List, Set, Type
+from typing import Callable, List, Type
 import os
 
 # Third Party
-from fms_acceleration.model_patcher import ModelPatcherRule, ModelPatcherTrigger
+from fms_acceleration.model_patcher import ModelPatcherTrigger
 from transformers import PretrainedConfig
 import torch
 
@@ -203,22 +203,6 @@ def trigger_fused_ops(
     return isinstance(module, attn_cls) and all(_is_loralayer(x) for x in _mods)
 
 
-# helper function to filter rules
-def filter_mp_rules(
-    rules: List[ModelPatcherRule],
-    filter_endswith: Set[str],
-    drop: bool = False,
-):
-    if drop:
-        # this means if any of the filter terms appear, we drop
-        return [
-            r for r in rules if not any(r.rule_id.endswith(x) for x in filter_endswith)
-        ]
-
-    # this means if any if the filter terms appear, we keep
-    return [r for r in rules if any(r.rule_id.endswith(x) for x in filter_endswith)]
-
-
 # helper function to get the hidden activation function str
 def get_hidden_activation_fn_key(config: PretrainedConfig):
     for key in KEY_HIDDEN_ACTIVATIONS:

plugins/fused-ops-and-kernels/src/fms_acceleration_foak/utils.py

@@ -12,8 +12,12 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+# Standard
+from typing import List, Set
+
 # Third Party
 from accelerate.utils import set_module_tensor_to_device
+from fms_acceleration.model_patcher import ModelPatcherRule
 from transformers.modeling_utils import is_fsdp_enabled
 import torch
 import torch.distributed as dist
@@ -74,3 +78,19 @@ def _all_reduce_hook(grad):
         # - this has to be done after all weight replacement happens
         A.weight.register_hook(_all_reduce_hook)
         B.weight.register_hook(_all_reduce_hook)
+
+
+# helper function to filter rules
+def filter_mp_rules(
+    rules: List[ModelPatcherRule],
+    filter_endswith: Set[str],
+    drop: bool = False,
+):
+    if drop:
+        # this means if any of the filter terms appear, we drop
+        return [
+            r for r in rules if not any(r.rule_id.endswith(x) for x in filter_endswith)
+        ]
+
+    # this means if any if the filter terms appear, we keep
+    return [r for r in rules if any(r.rule_id.endswith(x) for x in filter_endswith)]

plugins/fused-ops-and-kernels/tests/test_model_utils.py

@@ -2,7 +2,7 @@
 from fms_acceleration.model_patcher import ModelPatcherRule
 
 # First Party
-from fms_acceleration_foak.models.utils import filter_mp_rules
+from fms_acceleration_foak.utils import filter_mp_rules
 
 
 def test_filter_mp_rules():