Support rope scaling (#1391)

* support rope scaling * use rope scaling * update to use rope config * update config args * use allowlist for config to enforce hygeine * allow llama3 rope config * add unit test * documented allowed llama config keys * Update llmfoundry/models/mpt/modeling_mpt.py * Address comments 1 * Apply suggestions from code review Co-authored-by: Daniel King <[email protected]> * Apply suggestions from code review Co-authored-by: Daniel King <[email protected]> * use same codepath for all the hf rotary embeddings * fix * update * test WIP but fix get/pop * change the thing being popped * give up on testing hf --------- Co-authored-by: Daniel King <[email protected]>
mosaicml · Jul 24, 2024 · 70586c4 · 70586c4
1 parent af37ea0
commit 70586c4
Show file tree

Hide file tree

Showing 5 changed files with 111 additions and 35 deletions.
diff --git a/llmfoundry/models/mpt/configuration_mpt.py b/llmfoundry/models/mpt/configuration_mpt.py
@@ -306,6 +306,7 @@ def _validate_config(self) -> None:
             'no_scaling',
             'linear',
             'dynamic',
+            'llama3',
         ]:
             raise ValueError(
                 'If using hf implementation of rope, the type should be one of "no_scaling", "linear" or "dynamic".',

diff --git a/llmfoundry/models/mpt/modeling_mpt.py b/llmfoundry/models/mpt/modeling_mpt.py
@@ -49,10 +49,8 @@
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
 )
-from transformers.models.llama.modeling_llama import \
-    LlamaDynamicNTKScalingRotaryEmbedding as HFDynamicNTKScalingRotaryEmbedding
-from transformers.models.llama.modeling_llama import \
-    LlamaLinearScalingRotaryEmbedding as HFLinearScalingRotaryEmbedding
+from transformers.models.llama.modeling_llama import LlamaConfig
+from transformers.models.llama.modeling_llama import LlamaRotaryEmbedding
 from transformers.models.llama.modeling_llama import \
     LlamaRotaryEmbedding as HFRotaryEmbedding
 
@@ -88,14 +86,62 @@
 log = logging.getLogger(__name__)
 
 
+class InvalidConfigAccessError(KeyError):
+    pass
+
+
+_ALLOWED_LLAMA_CONFIG_KEYS = {
+    # These are the only config keys that are set and are safe to read from
+    'rope_scaling',
+    'rope_theta',
+    'max_position_embeddings',
+    'hidden_size',
+    'num_attention_heads',
+
+    # Not set but llama modeling code tries to read this attribute
+    'partial_rotary_factor',
+
+    # Benign transformers attributes needed for __init__
+    '_get_generation_defaults',
+    'label2id',
+    'id2label',
+    'torch_dtype',
+    'problem_type',
+    '__class__',
+}
+
+
+class PartialLlamaConfig(LlamaConfig):
+    """Holds the rope config for Llama models and throws.
+
+    an `InvalidConfigAccessError` if any other config elements are read. This
+    class is necessary because the `LlamaRotaryEmbedding` class takes a full
+    `LlamaConfig` now instead of the old keyword arguments.
+    """
+
+    def __getattribute__(self, key: str):
+        if key not in _ALLOWED_LLAMA_CONFIG_KEYS:
+            raise InvalidConfigAccessError(key)
+
+        return super().__getattribute__(key)
+
+    def __getitem__(self, key: str):
+        if key not in _ALLOWED_LLAMA_CONFIG_KEYS:
+            raise InvalidConfigAccessError(key)
+
+        return super().__getitem__(key)
+
+
 def gen_rotary_embedding(
-    rope_head_dim: int,
     rope_impl: str,
     rope_theta: int,
     rope_dail_config: dict,
     rope_hf_config: dict,
     max_seq_len: int,
+    d_model: int,
+    n_heads: int,
 ):
+    rope_head_dim = d_model // n_heads
     if rope_impl == 'dail':
         return DAILRotaryEmbedding(
             dim=rope_head_dim,
@@ -108,32 +154,21 @@ def gen_rotary_embedding(
             'cpu',  # FSDP does not materialize modules with meta buffers, hence device is set to cpu
         )
     elif rope_impl == 'hf':
+        llama_rope_config = {**rope_hf_config}
+        llama_rope_config['rope_type'] = llama_rope_config.pop('type')
+        if llama_rope_config['rope_type'] == 'no_scaling':
+            llama_rope_config['rope_type'] = 'default'
+        partial_llama_config = PartialLlamaConfig(
+            rope_scaling=llama_rope_config,
+            rope_theta=rope_theta,
+            max_position_embeddings=max_seq_len,
+            hidden_size=d_model,
+            num_attention_heads=n_heads,
+        )
         if rope_hf_config['type'] == 'no_scaling':
-            return HFRotaryEmbeddingFoundry(
-                rope_head_dim,
-                max_position_embeddings=max_seq_len,
-                base=rope_theta,
-                device=
-                'cpu',  # FSDP does not materialize modules with meta buffers, hence device is set to cpu
-            )
-        elif rope_hf_config['type'] == 'linear':
-            return HFLinearScalingRotaryEmbedding(
-                rope_head_dim,
-                max_position_embeddings=max_seq_len,
-                base=rope_theta,
-                scaling_factor=rope_hf_config['factor'],
-                device=
-                'cpu',  # FSDP does not materialize modules with meta buffers, hence device is set to cpu
-            )
-        elif rope_hf_config['type'] == 'dynamic':
-            return HFDynamicNTKScalingRotaryEmbedding(
-                rope_head_dim,
-                max_position_embeddings=max_seq_len,
-                base=rope_theta,
-                scaling_factor=rope_hf_config['factor'],
-                device=
-                'cpu',  # FSDP does not materialize modules with meta buffers, hence device is set to cpu
-            )
+            return HFRotaryEmbeddingFoundry(config=partial_llama_config)
+        elif rope_hf_config['type'] in {'llama3', 'linear', 'dynamic'}:
+            return LlamaRotaryEmbedding(config=partial_llama_config)
     raise ValueError('rope_impl needs to be either dail or hf')
 
 
@@ -399,12 +434,13 @@ def __init__(self, config: MPTConfig):
         if self.rope:
             self.rope_impl = config.attn_config['rope_impl']
             self.rotary_embedding = gen_rotary_embedding(
-                rope_head_dim=config.d_model // config.n_heads,
                 rope_impl=self.rope_impl,
                 rope_theta=config.attn_config['rope_theta'],
                 rope_dail_config=config.attn_config['rope_dail_config'],
                 rope_hf_config=config.attn_config['rope_hf_config'],
                 max_seq_len=self.config.max_seq_len,
+                d_model=config.d_model,
+                n_heads=config.n_heads,
             )
 
         if config.init_device != 'meta':

diff --git a/tests/models/layers/test_flash_torch.py b/tests/models/layers/test_flash_torch.py
@@ -251,12 +251,13 @@ def gen_bias(attn_impl: str):
         rotary_emb_w_meta_info = None
         if rope:
             rotary_embedding = gen_rotary_embedding(
-                rope_head_dim=cfg.d_model // cfg.n_heads,
                 rope_impl=pos_emb_config['rope_impl'],
                 rope_theta=pos_emb_config['rope_theta'],
                 rope_dail_config=pos_emb_config.get('rope_dail_config', {}),
                 rope_hf_config=pos_emb_config.get('rope_hf_config', {}),
                 max_seq_len=s,
+                d_model=cfg.d_model,
+                n_heads=cfg.n_heads,
             ).to(device)
             pos = torch.arange(s).unsqueeze(0).to(device=device)
             # adjust the position indices to account for padding tokens
@@ -664,12 +665,13 @@ def gen_bias(attn_impl: str):
         rotary_emb_w_meta_info = None
         if rope:
             rotary_embedding = gen_rotary_embedding(
-                rope_head_dim=cfg['d_model'] // cfg['n_heads'],
                 rope_impl=pos_emb_config['rope_impl'],
                 rope_theta=pos_emb_config['rope_theta'],
                 rope_dail_config=pos_emb_config.get('rope_dail_config', {}),
                 rope_hf_config=pos_emb_config.get('rope_hf_config', {}),
                 max_seq_len=s,
+                d_model=cfg['d_model'],
+                n_heads=cfg['n_heads'],
             ).to(device)
             pos = torch.arange(s).unsqueeze(0).to(device=device)
             # adjust the position indices to account for padding tokens

diff --git a/tests/models/test_rope_dail_vs_hf.py b/tests/models/test_rope_dail_vs_hf.py
@@ -77,12 +77,13 @@ def test_rope_dail_vs_hf(attn_type: str, seq_len: int, device: str = 'cuda'):
         }
 
         dail_rope = gen_rotary_embedding(
-            rope_head_dim=cfg.d_model // cfg.n_heads,
             rope_impl=dail_rope_config['rope_impl'],
             rope_theta=dail_rope_config['rope_theta'],
             rope_dail_config=dail_rope_config['rope_dail_config'],
             rope_hf_config={},
             max_seq_len=seq_len,
+            d_model=cfg.d_model,
+            n_heads=cfg.n_heads,
         ).to('cuda')
         dail_rope_w_meta_info = {
             'impl': 'dail',
@@ -92,12 +93,13 @@ def test_rope_dail_vs_hf(attn_type: str, seq_len: int, device: str = 'cuda'):
         }
 
         hf_rope = gen_rotary_embedding(
-            rope_head_dim=cfg.d_model // cfg.n_heads,
             rope_impl=hf_rope_config['rope_impl'],
             rope_theta=hf_rope_config['rope_theta'],
             rope_dail_config={},
             rope_hf_config=hf_rope_config['rope_hf_config'],
             max_seq_len=seq_len,
+            d_model=cfg.d_model,
+            n_heads=cfg.n_heads,
         ).to('cuda')
         pos = torch.arange(seq_len).unsqueeze(0).to(device='cuda')
         # adjust the position indices to account for padding tokens

diff --git a/tests/models/test_rope_scaling.py b/tests/models/test_rope_scaling.py
@@ -0,0 +1,35 @@
+# Copyright 2024 MosaicML LLM Foundry authors
+# SPDX-License-Identifier: Apache-2.0
+from transformers.models.llama.modeling_llama import LlamaRotaryEmbedding
+
+from llmfoundry.models.mpt.modeling_mpt import gen_rotary_embedding
+
+rope_config = {
+    'rope_theta': 500000.0,
+    'rope_impl': 'hf',
+    'rope_hf_config': {
+        'factor': 8.0,
+        'low_freq_factor': 1.0,
+        'high_freq_factor': 4.0,
+        'original_max_position_embeddings': 8192,
+        'type': 'llama3',
+    },
+}
+
+rope_dail_config = {}
+
+
+def test_rope_scaling():
+    d_model = 128
+    n_heads = 32
+    max_seq_len = 65536
+
+    embedding = gen_rotary_embedding(
+        d_model=d_model,
+        n_heads=n_heads,
+        rope_dail_config=rope_dail_config,
+        max_seq_len=max_seq_len,
+        **rope_config,
+    )
+
+    assert isinstance(embedding, LlamaRotaryEmbedding)