[REVIEW] Fix Padding Related Bugs: Crossfit (#66)

* Add crossfit bits Signed-off-by: Vibhu Jawa <[email protected]> * Add padding fixes Signed-off-by: Vibhu Jawa <[email protected]> * Fix test Signed-off-by: Vibhu Jawa <[email protected]> * Add docstrings Signed-off-by: Vibhu Jawa <[email protected]> * fix torch import Signed-off-by: Vibhu Jawa <[email protected]> * fix torch import Signed-off-by: Vibhu Jawa <[email protected]> * fix padding to only pad the last dim Signed-off-by: Vibhu Jawa <[email protected]> * fix padding tests Signed-off-by: Vibhu Jawa <[email protected]> * Add test for left/right Signed-off-by: Vibhu Jawa <[email protected]> * Skip test for cf_loader Signed-off-by: Vibhu Jawa <[email protected]> * Fix bugs in clipping Signed-off-by: Vibhu Jawa <[email protected]> * Fix bugs in clipping Signed-off-by: Vibhu Jawa <[email protected]> * Add early stopping to HF memory estimation Signed-off-by: Vibhu Jawa <[email protected]> * Fix copy-right year Signed-off-by: Vibhu Jawa <[email protected]> * Add copyright year Signed-off-by: Vibhu Jawa <[email protected]> * Address last of Ryan's reviews Signed-off-by: Vibhu Jawa <[email protected]> * Skip loading model if its allready fitted Signed-off-by: Vibhu Jawa <[email protected]> * Use self.load_cfg instead of AutoConfig.from_pretrained Signed-off-by: Vibhu Jawa <[email protected]> * Use self.load_cfg instead of AutoConfig.from_pretrained Signed-off-by: Vibhu Jawa <[email protected]> * Fix memory_curve_utils and skip loading cfg/tokenizer here Signed-off-by: Vibhu Jawa <[email protected]> --------- Signed-off-by: Vibhu Jawa <[email protected]>
rapidsai · Aug 5, 2024 · 0cc2993 · 0cc2993
1 parent 005e2fc
commit 0cc2993
Show file tree

Hide file tree

Showing 10 changed files with 615 additions and 98 deletions.
diff --git a/crossfit/__init__.py b/crossfit/__init__.py
@@ -84,6 +84,7 @@ def __call__(self, *args, **kwargs):
 load_dataset = LazyLoader("crossfit.dataset.load.load_dataset")
 embed = LazyLoader("crossfit.report.beir.embed.embed")
 beir_report = LazyLoader("crossfit.report.beir.report.beir_report")
+utils = LazyLoader("crossfit.utils")
 
 __all__.extend(
     [

diff --git a/crossfit/backend/torch/hf/memory_curve_utils.py b/crossfit/backend/torch/hf/memory_curve_utils.py
@@ -0,0 +1,89 @@
+# Copyright 2024 NVIDIA CORPORATION
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import gc
+
+import joblib
+import numpy as np
+import torch
+from sklearn.linear_model import LinearRegression
+from tqdm import tqdm
+from transformers import PreTrainedModel
+
+from crossfit.utils.model_adapter import adapt_model_input
+
+
+def fit_memory_estimate_curve(
+    model: PreTrainedModel,
+    path_or_name: str,
+    start_batch_size: int = 1,
+    end_batch_size: int = 2048,
+    batch_size_increment: int = 256,
+    start_seq_len: int = 1,
+    end_seq_len: int = 2048,
+    seq_len_increment: int = 64,
+    mem_model_path: str = None,
+) -> LinearRegression:
+    print(f"Fitting memory estimate curve for model: {path_or_name}")
+
+    device = next(model.parameters()).device
+    X: list[list[int]] = []
+    y: list[float] = []
+
+    batch_size_pbar = tqdm(
+        range(start_batch_size, end_batch_size + 1, batch_size_increment), desc="Batch size"
+    )
+    for batch_size in batch_size_pbar:
+        seq_len_pbar = tqdm(
+            range(start_seq_len, end_seq_len + 1, seq_len_increment),
+            desc="Sequence length",
+            leave=False,
+        )
+        for seq_len in seq_len_pbar:
+            torch.cuda.reset_peak_memory_stats()
+
+            batch = {
+                "input_ids": torch.randint(1, 501, (batch_size, seq_len)).to(device=device),
+                "attention_mask": torch.ones((batch_size, seq_len)).to(device=device),
+            }
+
+            try:
+                _ = adapt_model_input(model, batch)
+                memory_used = torch.cuda.max_memory_allocated() / (1024**2)  # Convert to MB
+                X.append([batch_size, seq_len, seq_len**2])
+                y.append(memory_used)
+
+            except RuntimeError as e:
+                if "out of memory" in str(e) or "out_of_memory" in str(e):
+                    # Early stopping for this batch size
+                    seq_len_pbar.close()
+                    break
+                else:
+                    raise e
+            finally:
+                del batch
+                if "outputs" in vars():
+                    del outputs
+                gc.collect()
+                torch.cuda.empty_cache()
+
+        # Check if we've hit the memory limit for all sequence lengths
+        if seq_len == start_seq_len:
+            batch_size_pbar.close()
+            break
+
+    mem_model = LinearRegression().fit(np.array(X), np.array(y))
+    joblib.dump(mem_model, mem_model_path)
+
+    return mem_model
diff --git a/crossfit/backend/torch/hf/model.py b/crossfit/backend/torch/hf/model.py
@@ -19,24 +19,65 @@
 import joblib
 import numpy as np
 import torch
-from sklearn.linear_model import LinearRegression
-from tqdm import tqdm
 from transformers import AutoConfig, AutoModel, AutoTokenizer
 
+from crossfit.backend.torch.hf.memory_curve_utils import fit_memory_estimate_curve
 from crossfit.backend.torch.model import Model
 from crossfit.dataset.home import CF_HOME
-from crossfit.utils.model_adapter import adapt_model_input
 
 
 class HFModel(Model):
-    def __init__(self, path_or_name: str, max_mem_gb: int = 16, training=False):
+    def __init__(
+        self,
+        path_or_name: str,
+        max_mem_gb: int = 16,
+        training: bool = False,
+        start_batch_size: int = 1,
+        end_batch_size: int = 2048,
+        batch_size_increment: int = 256,
+        start_seq_len: int = 1,
+        seq_len_increment: int = 64,
+    ):
         super().__init__(path_or_name, max_mem_gb)
+        self.start_batch_size = start_batch_size
+        self.end_batch_size = end_batch_size
+        self.batch_size_increment = batch_size_increment
+        self.start_seq_len = start_seq_len
+        self.seq_len_increment = seq_len_increment
 
-        if not training:
-            with torch.no_grad():
-                self.fit_memory_estimate_curve()
+        cache_dir = os.path.join(CF_HOME, "memory", self.load_cfg()._name_or_path)
+        os.makedirs(cache_dir, exist_ok=True)
+        mem_model_path = os.path.join(cache_dir, "mem_model.pkl")
+        if os.path.exists(mem_model_path):
+            self.mem = joblib.load(mem_model_path)
         else:
-            self.fit_memory_estimate_curve()
+            model = self.load_model("cuda") if not training else None
+            end_seq_len = self.max_seq_length()
+            if not training:
+                with torch.no_grad():
+                    self.mem = fit_memory_estimate_curve(
+                        model=model,
+                        path_or_name=self.path_or_name,
+                        start_batch_size=start_batch_size,
+                        end_batch_size=end_batch_size,
+                        start_seq_len=start_seq_len,
+                        end_seq_len=end_seq_len,
+                        batch_size_increment=batch_size_increment,
+                        seq_len_increment=seq_len_increment,
+                        mem_model_path=mem_model_path,
+                    )
+            else:
+                self.mem = fit_memory_estimate_curve(
+                    model=model,
+                    path_or_name=self.path_or_name,
+                    start_batch_size=start_batch_size,
+                    end_batch_size=end_batch_size,
+                    start_seq_len=start_seq_len,
+                    end_seq_len=end_seq_len,
+                    batch_size_increment=batch_size_increment,
+                    seq_len_increment=seq_len_increment,
+                    mem_model_path=mem_model_path,
+                )
 
     def load_on_worker(self, worker, device="cuda"):
         worker.torch_model = self.load_model(device)
@@ -60,77 +101,19 @@ def load_tokenizer(self):
     def load_cfg(self):
         return AutoConfig.from_pretrained(self.path_or_name)
 
-    def fit_memory_estimate_curve(self, model=None):
-        remove_model = False
-        if model is None:
-            remove_model = True
-            model = self.load_model(device="cuda")
-
-        cache_dir = os.path.join(CF_HOME, "memory", self.load_cfg()._name_or_path)
-        mem_model_path = os.path.join(cache_dir, "mem_model.pkl")
-
-        if os.path.exists(mem_model_path):
-            self.mem = joblib.load(mem_model_path)
-
-            return self
-
-        print(f"Fitting memory estimate curve for model: {self.path_or_name}")
-
-        device = next(model.parameters()).device
-        X = []
-        y = []
-
-        max_seq = self.max_seq_length()
-        for batch_size in tqdm(range(2048, 0, -256)):
-            if batch_size <= 0:
-                continue
-
-            for seq_len in range(max_seq, 0, -64):
-                if seq_len <= 0:
-                    continue
-
-                torch.cuda.reset_peak_memory_stats()
-
-                batch = {
-                    "input_ids": torch.randint(1, 501, (batch_size, seq_len)).to(device=device),
-                    "attention_mask": torch.ones((batch_size, seq_len)).to(device=device),
-                }
-
-                try:
-                    _ = adapt_model_input(model, batch)
-                    memory_used = torch.cuda.max_memory_allocated() / (1024**2)  # Convert to MB
-                    X.append([batch_size, seq_len, seq_len**2])
-                    y.append(memory_used)
-
-                except RuntimeError as e:
-                    if "out of memory" in str(e) or "out_of_memory" in str(e):
-                        pass
-                    else:
-                        raise e
-                finally:
-                    del batch
-                    if "outputs" in vars():
-                        del outputs
-                    gc.collect()
-                    torch.cuda.empty_cache()
-
-        self.mem = LinearRegression().fit(np.array(X), np.array(y))
-        os.makedirs(cache_dir, exist_ok=True)
-        joblib.dump(self.mem, mem_model_path)
-
-        if remove_model:
-            del model
-        gc.collect()
-        torch.cuda.empty_cache()
-
     def estimate_memory(self, max_num_tokens: int, batch_size: int) -> int:
         predicted_memory = self.mem.predict(
             np.array([[batch_size, max_num_tokens, max_num_tokens**2]])
         )
         return predicted_memory[0] / 1024  # Convert from MB to GB
 
     def max_seq_length(self) -> int:
-        return self.load_cfg().max_position_embeddings
+        max_seq_length = self.load_tokenizer().model_max_length
+        # Guard against the HF bug
+        # which sets max_seq_length to max(int) for some models
+        if max_seq_length > 1e5:
+            max_seq_length = self.load_cfg().max_position_embeddings
+        return max_seq_length
 
 
 class SentenceTransformerModel(HFModel):

diff --git a/crossfit/backend/torch/loader.py b/crossfit/backend/torch/loader.py
@@ -1,4 +1,4 @@
-# Copyright 2023 NVIDIA CORPORATION
+# Copyright 2024 NVIDIA CORPORATION
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -22,6 +22,7 @@
 from crossfit.data.array.conversion import convert_array
 from crossfit.data.array.dispatch import crossarray
 from crossfit.data.dataframe.dispatch import CrossFrame
+from crossfit.op.tokenize import clip_tokens
 from crossfit.utils.model_adapter import adapt_model_input
 
 DEFAULT_BATCH_SIZE = 512
@@ -36,7 +37,14 @@ def __init__(self, data: Dict[str, torch.Tensor], batch_size: int, progress_bar=
     def __init__(self, data: CrossFrame, batch_size: int, progress_bar=None):
         ...
 
-    def __init__(self, data, batch_size: int, progress_bar=None, max_seq_len=None):
+    def __init__(
+        self,
+        data,
+        batch_size: int,
+        progress_bar=None,
+        max_seq_len=None,
+        padding_side: str = "right",
+    ):
         self.data = CrossFrame(data).cast(torch.Tensor)
         self.tensor_dict = self.data.to_dict()
         self._batch_size = batch_size
@@ -45,6 +53,7 @@ def __init__(self, data, batch_size: int, progress_bar=None, max_seq_len=None):
         self._to_map = []
         self.progress_bar = progress_bar
         self.max_seq_len = max_seq_len
+        self.padding_side = padding_side
 
     def map(self, fn):
         self._to_map.append(fn)
@@ -66,7 +75,10 @@ def __next__(self):
 
         batch = {key: val[self.current_idx : end] for key, val in self.tensor_dict.items()}
         if self.max_seq_len is not None:
-            batch = {key: val[:, : self.max_seq_len] for key, val in batch.items()}
+            if self.padding_side == "right":
+                batch = {key: val[:, : self.max_seq_len] for key, val in batch.items()}
+            else:
+                batch = {key: val[:, -self.max_seq_len :] for key, val in batch.items()}
 
         self.current_idx += self.batch_size
 
@@ -97,14 +109,27 @@ def __init__(
         self.to_ignore = to_ignore or []
         self.to_ignore.append("seq_length")
         self.model = model
+        tokenizer = self.model.load_tokenizer()
+        pad_token_id = tokenizer.pad_token_id
+        padding_side = tokenizer.padding_side
 
+        if padding_side not in ["right", "left"]:
+            raise ValueError("padding_side must be either 'right' or 'left'")
+
+        self.pad_token_id = pad_token_id
         frame = CrossFrame(data).cast(torch.Tensor)
-        seq_length = (frame[sort_key] != 0).sum(axis=1)
+        seq_length = (frame[sort_key] != self.pad_token_id).sum(axis=1)
         self.sorted_indices = seq_length.argsort(descending=True)
         frame = frame.apply(lambda x: x[self.sorted_indices])
         frame = frame.assign(seq_length=seq_length[self.sorted_indices])
 
-        super().__init__(frame, initial_batch_size, progress_bar=progress_bar)
+        super().__init__(
+            frame,
+            initial_batch_size,
+            progress_bar=progress_bar,
+            max_seq_len=self.model.max_seq_length(),
+            padding_side=padding_side,
+        )
         self.splits = self._find_optimal_splits()
 
     def sort_column(self, col):
@@ -128,8 +153,6 @@ def __next__(self):
         else:
             start = self.splits[self.current_idx - 1]
 
-        _tokens = self.tensor_dict["seq_length"]
-
         end = min(self.splits[self.current_idx], self.num_rows)
         while end > start:
             try:
@@ -138,8 +161,13 @@ def __next__(self):
                     for key, val in self.tensor_dict.items()
                     if key not in self.to_ignore
                 }
-                clip_len = min(max(_tokens[start], _tokens[end - 1]), self.model.max_seq_length())
-                batch = {key: val[:, :clip_len] for key, val in batch.items()}
+                batch = clip_tokens(
+                    token_o=batch,
+                    max_length=self.max_seq_len,
+                    padding_side=self.padding_side,
+                    pad_token_id=self.pad_token_id,
+                    return_type="pt",
+                )
 
                 for fn in self._to_map:
                     batch = adapt_model_input(fn, batch)

diff --git a/crossfit/backend/torch/op/base.py b/crossfit/backend/torch/op/base.py
@@ -26,6 +26,7 @@
 from crossfit.backend.torch.loader import DEFAULT_BATCH_SIZE, InMemoryLoader, SortedSeqLoader
 from crossfit.backend.torch.model import Model
 from crossfit.op.base import Op
+from crossfit.utils.torch_utils import concat_and_pad_tensors
 
 
 class Predictor(Op):
@@ -66,6 +67,7 @@ def call(self, data, partition_info=None):
             loader = InMemoryLoader(
                 data[["input_ids", "attention_mask"]],
                 batch_size=self.batch_size,
+                padding_side=self.model.load_tokenizer().padding_side,
                 progress_bar=self.create_progress_bar(len(data), partition_info),
                 max_seq_len=self.model.max_seq_length(),
             )
@@ -83,7 +85,11 @@ def call(self, data, partition_info=None):
             all_outputs_ls.append(output)
 
         out = cudf.DataFrame(index=index)
-        outputs = cp.asarray(torch.cat(all_outputs_ls, dim=0))
+        outputs = cp.asarray(
+            concat_and_pad_tensors(
+                all_outputs_ls, pad_token_id=loader.pad_token_id, padding_side=loader.padding_side
+            )
+        )
         _index = loader.sort_column(index.values) if self.sorted_data_loader else index
         if len(outputs.shape) <= 2:
             out[self.pred_output_col] = create_list_series_from_1d_or_2d_ar(outputs, _index)