[whisper] compile compatibility with long-form decoding

docs/source/en/model_doc/whisper.md

@@ -72,7 +72,7 @@ Here is a step-by-step guide to transcribing an audio sample using a pre-trained
 ' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
 ```
 
-Whisper is compatible with the following optimisations:
+Whisper is compatible with the following optimisations for both short and long-form generation:
 - [PyTorch Scaled Dot Product Attention (SDPA)](../perf_infer_gpu_one#pytorch-scaled-dot-product-attention): flash attention and memory-efficient attention kernels. Enabled by default for `torch>=2.1.1`.
 - [Flash Attention 2](../perf_infer_gpu_one#flashattention-2): improved implementation of flash attention through better parallelism and work partitioning. 
 - [torch.compile](../llm_optims#static-kv-cache-and-torchcompile): JIT-compile the forward pass to dispatch to efficient fused kernels.
@@ -101,7 +101,8 @@ As an example, the following codesnippet enables SDPA and `torch.compile` for up
 ... ).input_features
 
 >>> # Compile the forward pass
->>> _ = model.generate(input_features)
+>>> for _ in range(2):
+>>>     model.generate(input_features)
 
 >>> # Generate token ids using compiled graph (fast!)
 >>> predicted_ids = model.generate(input_features)

src/transformers/models/whisper/generation_whisper.py

@@ -126,12 +126,24 @@ def _get_attr_from_logit_processors(logits_processor, logit_processor_class, att
 
 
 def _pad_to_max_length(
-    current_segments, pad_token_id, device, padding="right", bos_token_tensor=None, cut_off_length=None
+    current_segments,
+    pad_token_id,
+    device,
+    padding_side="right",
+    padding="longest",
+    bos_token_tensor=None,
+    cut_off_length=None,
 ):
     max_total_length = 0
     sequences = []
-    if padding not in ["right", "left"]:
-        raise ValueError(f"`padding` must be either 'right' or 'left', not {padding}")
+
+    if padding_side not in ["right", "left"]:
+        raise ValueError(f"`padding_side` must be either 'right' or 'left', not {padding_side}")
+
+    if padding not in ["longest", "max_length"]:
+        raise ValueError(f"`padding` must be either 'longest' or 'max_length', not {padding}")
+    elif padding == "max_length" and cut_off_length is None:
+        raise ValueError("`cut_off_length` must be specified when `padding='max_length'`")
 
     for current_segment_list in current_segments:
         if current_segment_list is not None and len([d["tokens"] for d in current_segment_list]) > 0:
@@ -150,9 +162,10 @@ def _pad_to_max_length(
         else:
             sequences.append(torch.tensor([], device=device))
 
+    max_total_length = cut_off_length + 1 if padding == "max_length" else max_total_length
     for i in range(len(current_segments)):
         pad_length = max_total_length - len(sequences[i])
-        pad = (0, pad_length) if padding == "right" else (pad_length, 0)
+        pad = (0, pad_length) if padding_side == "right" else (pad_length, 0)
         sequences[i] = F.pad(sequences[i], pad=pad, value=pad_token_id)
 
     sequences = torch.stack(sequences, dim=0)
@@ -672,6 +685,7 @@ def generate(
                 return_token_timestamps=return_token_timestamps,
                 do_condition_on_prev_tokens=do_condition_on_prev_tokens,
                 is_shortform=is_shortform,
+                batch_size=batch_size,
                 kwargs=kwargs,
             )
 
@@ -712,7 +726,7 @@ def generate(
         )
 
         sequences = _pad_to_max_length(
-            final_segments, generation_config.pad_token_id, device=self.device, padding="right"
+            final_segments, generation_config.pad_token_id, device=self.device, padding_side="right"
         )
 
         # 8. If we return all segments, the predicted output sequences are put under `"sequences"`.
@@ -775,6 +789,7 @@ def generate_with_fallback(
         return_token_timestamps,
         do_condition_on_prev_tokens,
         is_shortform,
+        batch_size,
         kwargs,
     ):
         kwargs = copy.copy(kwargs)
@@ -798,6 +813,22 @@ def generate_with_fallback(
             for key in ["do_sample", "temperature", "num_beams"]:
                 if key in generate_kwargs:
                     del generate_kwargs[key]
+
+            cur_bsz = decoder_input_ids.shape[0]
+            if generation_config.cache_implementation == "static" and cur_bsz < batch_size:
+                segment_input = F.pad(segment_input, (0, 0, 0, 0, 0, batch_size - cur_bsz), value=0)
+                decoder_input_ids = F.pad(
+                    decoder_input_ids, (0, 0, 0, batch_size - cur_bsz), value=generation_config.pad_token_id
+                )
+                if generate_kwargs.get("decoder_attention_mask") is not None:
+                    generate_kwargs["decoder_attention_mask"] = F.pad(
+                        generate_kwargs["decoder_attention_mask"], (0, 0, 0, batch_size - cur_bsz), value=True
+                    )
+                if generate_kwargs.get("encoder_outputs") is not None:
+                    generate_kwargs["encoder_outputs"] = F.pad(
+                        generate_kwargs["encoder_outputs"], (0, 0, 0, 0, 0, batch_size - cur_bsz), value=0
+                    )
+
             seek_outputs = super().generate(
                 segment_input,
                 generation_config=generation_config,
@@ -820,6 +851,10 @@ def generate_with_fallback(
                 is_shortform=is_shortform,
             )
 
+            if cur_bsz < batch_size:
+                seek_sequences = seek_sequences[:cur_bsz]
+                seek_outputs = seek_outputs[:cur_bsz]
+
             # 6.7 Extract cut sequences from every sequence and check if fallback should be applied
             # Loop over each decoded audio individually as each decoding can be of a different length
             new_fallback_index_map = []
@@ -925,17 +960,27 @@ def split_by_batch_index(values, key, batch_idx, is_shortform):
                 if not is_shortform:
                     # we don't save `past_key_values` as this is too costly for longform
                     return None
+                elif isinstance(values, EncoderDecoderCache):
+                    all_past_key_values = []
+                    for layer_idx in range(self.config.decoder_layers):
+                        layer_past_key_values = []
+                        for cache_cls in [values.self_attention_cache, values.cross_attention_cache]:
+                            for v in [cache_cls.key_cache, cache_cls.value_cache]:
+                                layer_past_key_values.append(v[layer_idx][batch_idx][None].cpu())
+                        all_past_key_values.append(tuple(layer_past_key_values))
+                    return tuple(all_past_key_values)
                 else:
-                    return tuple(tuple(w[batch_idx][None].cpu() for w in values[v]) for v in range(len(values)))
+                    all_past_key_values = []
+                    for v in range(len(values)):
+                        layer_past_key_values = []
+                        for w in values[v]:
+                            layer_past_key_values.append(w[batch_idx][None].cpu())
+                        all_past_key_values.append(tuple(layer_past_key_values))
+                    return tuple(all_past_key_values)
 
             return values[batch_idx].cpu()
 
         sequence_tokens = seek_outputs["sequences"]
-
-        if hasattr(seek_outputs, "past_key_values") and seek_outputs.past_key_values is not None:
-            if isinstance(seek_outputs["past_key_values"], EncoderDecoderCache):
-                seek_outputs.past_key_values = seek_outputs.past_key_values.to_legacy_cache()
-
         seek_outputs = [
             {k: split_by_batch_index(v, k, i, is_shortform) for k, v in seek_outputs.items()}
             for i in range(sequence_tokens.shape[0])
@@ -1613,11 +1658,14 @@ def _prepare_decoder_input_ids(
                 one_tensor = torch.ones((cur_bsz, 1), device=device, dtype=torch.long)
                 prev_ids = prev_start_of_text * one_tensor[0] if prev_start_of_text is not None else None
 
+            padding = "max_length" if generation_config.cache_implementation == "static" else "longest"
+
             prev_tokens = _pad_to_max_length(
                 active_segments,
                 generation_config.pad_token_id,
                 device=device,
-                padding="left",
+                padding_side="left",
+                padding=padding,
                 bos_token_tensor=prev_ids,
                 cut_off_length=cut_off_length,
             )

src/transformers/models/whisper/modeling_whisper.py

@@ -1835,8 +1835,10 @@ def prepare_inputs_for_generation(
 
             decoder_input_ids = decoder_input_ids[:, remove_prefix_length:]
 
-            if decoder_position_ids is not None and decoder_position_ids.shape[1] > decoder_input_ids.shape[1]:
+            if decoder_position_ids is not None:
                 decoder_position_ids = decoder_position_ids[:, remove_prefix_length:]
+                # This `clone` call is needed to avoid recapturing cuda graphs with `torch.compile`'s  `mode="reduce-overhead`, as otherwise the input `position_ids` would have various stride during the decoding. Here, simply using `.contiguous()` is not sufficient as in the batch size = 1 case, `position_ids` is already contiguous but with varying stride which retriggers a capture.
+                decoder_position_ids = decoder_position_ids.clone(memory_format=torch.contiguous_format)
 
         if cache_position is None:
             cache_position = torch.arange(
@@ -1845,6 +1847,36 @@ def prepare_inputs_for_generation(
         elif use_cache:
             cache_position = cache_position[-decoder_input_ids.shape[1] :]
 
+        # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
+        # recompiles graphs as the stride of the inputs is a guard. Ref: https://github.com/huggingface/transformers/pull/29114
+        decoder_input_ids = decoder_input_ids.contiguous()
+
+        if (
+            isinstance(past_key_values, EncoderDecoderCache)
+            and (
+                isinstance(past_key_values.self_attention_cache, StaticCache)
+                or isinstance(past_key_values.cross_attention_cache, StaticCache)
+            )
+            and decoder_attention_mask is not None
+            and decoder_attention_mask.ndim == 2
+        ):
+            batch_size, sequence_length = decoder_input_ids.shape
+            device = decoder_input_ids.device
+
+            dtype = self.proj_out.weight.dtype
+            min_dtype = torch.finfo(dtype).min
+
+            decoder_attention_mask = _prepare_4d_causal_attention_mask_with_cache_position(
+                decoder_attention_mask,
+                sequence_length=sequence_length,
+                target_length=past_key_values.self_attention_cache.get_max_length(),
+                dtype=dtype,
+                device=device,
+                min_dtype=min_dtype,
+                cache_position=cache_position,
+                batch_size=batch_size,
+            )
+
         return {
             "encoder_outputs": encoder_outputs,
             "past_key_values": past_key_values,

tests/models/whisper/test_modeling_whisper.py

@@ -3386,6 +3386,66 @@ def test_tiny_static_generation(self):
         # assert re-ordered generations match those from eager
         assert (eager_generated_ids[permutation_idx, :] == static_generated_ids).all()
 
+    @slow
+    def test_tiny_static_generation_long_form(self):
+        import torch._dynamo.config
+
+        # only permit 4 compilations: 2 prefill steps and 2 decoding steps (1 for each of conditioned/not conditioned)
+        torch._dynamo.config.cache_size_limit = 4
+
+        processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+        model.to(torch_device)
+
+        dataset = load_dataset("distil-whisper/meanwhile", "default")["test"]
+        dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
+        input_speech = [audio["array"] for audio in dataset[2:4]["audio"]]
+
+        inputs = processor(
+            input_speech,
+            return_tensors="pt",
+            padding="longest",
+            truncation=False,
+            return_attention_mask=True,
+            sampling_rate=16_000,
+        )
+        inputs = inputs.to(torch_device)
+
+        gen_kwargs = {
+            "return_timestamps": True,
+            "no_speech_threshold": 0.6,
+            "temperature": (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
+            "compression_ratio_threshold": 1.35,
+            "condition_on_prev_tokens": True,  # conditioning on prev tokens introduces a recompile on the second time step
+            "logprob_threshold": -1.0,
+            "num_beams": 1,
+        }
+
+        set_seed(42)
+        eager_generated_ids = model.generate(**inputs, **gen_kwargs)
+
+        # compile the forward pass and assert equivalence
+        model.generation_config.cache_implementation = "static"
+        model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
+
+        set_seed(42)
+        static_generated_ids = model.generate(**inputs, **gen_kwargs)
+        assert (eager_generated_ids == static_generated_ids).all()
+
+        # check the compiled graph can be re-used and that the cache is correctly reset
+        # reverse the ordering of the input features
+        input_features = inputs.input_features
+        permutation_idx = (
+            torch.arange(input_features.shape[0], 0, step=-1, dtype=torch.long, device=input_features.device) - 1
+        )
+        input_features = input_features[permutation_idx, ...]
+        attention_mask = inputs.attention_mask[permutation_idx, ...]
+
+        set_seed(42)
+        static_generated_ids = model.generate(input_features, attention_mask=attention_mask, **gen_kwargs)
+        # assert re-ordered generations match those from eager
+        assert (eager_generated_ids[permutation_idx, :] == static_generated_ids).all()
+
 
 def prepare_whisper_encoder_inputs_dict(config, input_features, head_mask=None):
     if head_mask is None: