Fix Falcon ONNX export with alibi (#1524)

fix falcon export with alibi

optimum/exporters/onnx/model_patcher.py

@@ -15,12 +15,12 @@
 import dataclasses
 import functools
 import inspect
+import math
 import types
 from typing import TYPE_CHECKING, Any, Callable, Dict, Optional, Tuple, Union
 
 from packaging import version
 from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions
-from transformers.models.falcon.modeling_falcon import build_alibi_tensor
 from transformers.models.speecht5.modeling_speecht5 import SpeechT5EncoderWithSpeechPrenet
 from transformers.utils import is_torch_available
 
@@ -310,7 +310,9 @@ def falcon_model_forward_without_kv_reformatting(
         attention_mask = attention_mask.to(hidden_states.device)
 
     if self.use_alibi:
-        alibi = build_alibi_tensor(attention_mask, self.num_heads, dtype=hidden_states.dtype)
+        # NOTE: we use a patched build_alibi_tensor.
+        alibi = falcon_build_alibi_tensor_patched(attention_mask, self.num_heads, dtype=hidden_states.dtype)
+        # alibi = build_alibi_tensor(attention_mask, self.num_heads, dtype=hidden_states.dtype)
     else:
         alibi = None
         if position_ids is None:
@@ -431,6 +433,37 @@ def __init__(
             self.original_make_causal = AttentionMaskConverter._make_causal_mask
 
 
+def falcon_build_alibi_tensor_patched(
+    attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype
+) -> torch.Tensor:
+    batch_size, seq_length = attention_mask.shape
+    closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
+    base = torch.tensor(
+        2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
+    )
+    powers = torch.arange(1, 1 + closest_power_of_2, device=attention_mask.device, dtype=torch.int32)
+    slopes = torch.pow(base, powers)
+
+    if closest_power_of_2 != num_heads:
+        extra_base = torch.tensor(
+            2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
+        )
+        num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
+        extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=attention_mask.device, dtype=torch.int32)
+        slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
+
+    # Note: alibi will added to the attention bias that will be applied to the query, key product of attention
+    # => therefore alibi will have to be of shape (batch_size, num_heads, query_length, key_length)
+    # => here we set (batch_size=1, num_heads=num_heads, query_length=1, key_length=max_length)
+    # => the query_length dimension will then be broadcasted correctly
+    # This is more or less identical to T5's relative position bias:
+    # https://github.com/huggingface/transformers/blob/f681437203baa7671de3174b0fa583c349d9d5e1/src/transformers/models/t5/modeling_t5.py#L527
+    # NOTE: remove the .bfloat16() cast here as PyTorch ONNX export rather casts to complex128 if this is used, resulting in a onnxruntime.capi.onnxruntime_pybind11_state.InvalidGraph error.
+    arange_tensor = ((attention_mask.cumsum(dim=-1) - 1) * attention_mask)[:, None, :]
+    alibi = slopes[..., None] * arange_tensor
+    return alibi.reshape(batch_size * num_heads, 1, seq_length).to(dtype)
+
+
 class FalconModelPatcher(DecoderModelPatcher):
     def __enter__(self):
         super().__enter__()

tests/exporters/exporters_utils.py

@@ -72,7 +72,17 @@
         ],
         "mohitsha/tiny-random-testing-bert2gpt2": ["text2text-generation", "text2text-generation-with-past"],
     },
-    "falcon": "fxmarty/really-tiny-falcon-testing",
+    "falcon": {
+        "fxmarty/really-tiny-falcon-testing": [
+            "feature-extraction",
+            "feature-extraction-with-past",
+            "question-answering",
+            "text-generation",
+            "text-generation-with-past",
+            "token-classification",
+        ],
+        "fxmarty/tiny-testing-falcon-alibi": ["text-generation", "text-generation-with-past"],
+    },
     "flaubert": "hf-internal-testing/tiny-random-flaubert",
     "gpt2": "hf-internal-testing/tiny-random-gpt2",
     "gpt-bigcode": "hf-internal-testing/tiny-random-GPTBigCodeModel",