TF: GPT-J compatible with XLA generation (#17986)

src/transformers/models/gptj/modeling_tf_gptj.py

@@ -60,14 +60,12 @@ GPTJ_PRETRAINED_MODEL_ARCHIVE_LIST = [
 ]
 
 
-def fixed_pos_embedding(x: tf.Tensor, seq_dim: int = 1, seq_len: Optional[int] = None) -> Tuple[tf.Tensor, tf.Tensor]:
-    dim = shape_list(x)[-1]
-    if seq_len is None:
-        seq_len = shape_list(x)[seq_dim]
+def create_sinusoidal_positions(num_pos: int, dim: int) -> tf.Tensor:
     inv_freq = tf.cast(1.0 / (10000 ** (tf.range(0, dim, 2) / dim)), tf.float32)
-    seq_len_range = tf.cast(tf.range(seq_len), tf.float32)
-    sinusoid_inp = tf.cast(tf.einsum("i , j -> i j", seq_len_range, inv_freq), tf.float32)
-    return tf.cast(tf.sin(sinusoid_inp), dtype=x.dtype), tf.cast(tf.cos(sinusoid_inp), dtype=x.dtype)
+    sinusoid_inp = tf.cast(tf.einsum("i , j -> i j", tf.range(num_pos, dtype=tf.float32), inv_freq), tf.float32)
+    sin, cos = tf.sin(sinusoid_inp), tf.cos(sinusoid_inp)
+    out = tf.concat((sin, cos), axis=1)
+    return out
 
 
 def rotate_every_two(x: tf.Tensor) -> tf.Tensor:
@@ -77,11 +75,11 @@ def rotate_every_two(x: tf.Tensor) -> tf.Tensor:
     return rotate_half_tensor
 
 
-def apply_rotary_pos_emb(x: tf.Tensor, sincos: tf.Tensor, offset: int = 0) -> tf.Tensor:
+def apply_rotary_pos_emb(tensor: tf.Tensor, sincos: tf.Tensor) -> tf.Tensor:
     sin_pos, cos_pos = sincos
-    sin_pos = tf.repeat(sin_pos[None, offset : shape_list(x)[1] + offset, None, :], 2, 3)
-    cos_pos = tf.repeat(cos_pos[None, offset : shape_list(x)[1] + offset, None, :], 2, 3)
-    return (x * cos_pos) + (rotate_every_two(x) * sin_pos)
+    sin_pos = tf.repeat(sin_pos[:, :, None, :], 2, 3)
+    cos_pos = tf.repeat(cos_pos[:, :, None, :], 2, 3)
+    return (tensor * cos_pos) + (rotate_every_two(tensor) * sin_pos)
 
 
 class TFGPTJAttention(tf.keras.layers.Layer):
@@ -132,6 +130,8 @@ class TFGPTJAttention(tf.keras.layers.Layer):
             tf.cast(tf.experimental.numpy.tril(tf.ones((self.max_positions, self.max_positions))), tf.int8),
             (1, 1, self.max_positions, self.max_positions),
         )
+        pos_embd_dim = self.rotary_dim or self.embed_dim
+        self.embed_positions = create_sinusoidal_positions(self.max_positions, pos_embd_dim)
 
     def get_causal_mask(self, key_length, query_length) -> tf.Tensor:
         return tf.cast(self.lower_triangle_mask[:, :, key_length - query_length : key_length, :key_length], tf.bool)
@@ -207,8 +207,9 @@ class TFGPTJAttention(tf.keras.layers.Layer):
     def call(
         self,
         hidden_states: tf.Tensor,
-        attention_mask: Optional[tf.Tensor] = None,
         layer_past: Optional[Tuple[tf.Tensor, tf.Tensor]] = None,
+        attention_mask: Optional[tf.Tensor] = None,
+        position_ids: Optional[tf.Tensor] = None,
         head_mask: Optional[tf.Tensor] = None,
         use_cache: bool = False,
         output_attentions: bool = False,
@@ -221,13 +222,8 @@ class TFGPTJAttention(tf.keras.layers.Layer):
         key = self._split_heads(key, True)
         value = self._split_heads(value, False)
 
-        seq_len = shape_list(key)[1]
-        offset = 0
-
-        if layer_past is not None:
-            offset = shape_list(layer_past[0])[-2]
-            seq_len += offset
-
+        sincos = tf.gather(self.embed_positions, position_ids, axis=0)
+        sincos = tf.split(sincos, 2, axis=-1)
         if self.rotary_dim is not None:
             k_rot = key[:, :, :, : self.rotary_dim]
             k_pass = key[:, :, :, self.rotary_dim :]
@@ -235,16 +231,14 @@ class TFGPTJAttention(tf.keras.layers.Layer):
             q_rot = query[:, :, :, : self.rotary_dim]
             q_pass = query[:, :, :, self.rotary_dim :]
 
-            sincos = fixed_pos_embedding(k_rot, 1, seq_len=seq_len)
-            k_rot = apply_rotary_pos_emb(k_rot, sincos, offset=offset)
-            q_rot = apply_rotary_pos_emb(q_rot, sincos, offset=offset)
+            k_rot = apply_rotary_pos_emb(k_rot, sincos)
+            q_rot = apply_rotary_pos_emb(q_rot, sincos)
 
             key = tf.concat((k_rot, k_pass), axis=-1)
             query = tf.concat((q_rot, q_pass), axis=-1)
         else:
-            sincos = fixed_pos_embedding(key, 1, seq_len=seq_len)
-            key = apply_rotary_pos_emb(key, sincos, offset=offset)
-            query = apply_rotary_pos_emb(query, sincos, offset=offset)
+            key = apply_rotary_pos_emb(key, sincos)
+            query = apply_rotary_pos_emb(query, sincos)
 
         key = tf.transpose(key, (0, 2, 1, 3))
         query = tf.transpose(query, (0, 2, 1, 3))
@@ -310,6 +304,7 @@ class TFGPTJBlock(tf.keras.layers.Layer):
         hidden_states: tf.Tensor,
         layer_past: Optional[tf.Tensor] = None,
         attention_mask: Optional[tf.Tensor] = None,
+        position_ids: Optional[tf.Tensor] = None,
         head_mask: Optional[tf.Tensor] = None,
         use_cache: bool = False,
         output_attentions: bool = False,
@@ -317,9 +312,10 @@ class TFGPTJBlock(tf.keras.layers.Layer):
         residual = hidden_states
         hidden_states = self.ln_1(hidden_states)
         attn_outputs = self.attn(
-            hidden_states,
+            hidden_states=hidden_states,
             layer_past=layer_past,
             attention_mask=attention_mask,
+            position_ids=position_ids,
             head_mask=head_mask,
             use_cache=use_cache,
             output_attentions=output_attentions,
@@ -466,12 +462,13 @@ class TFGPTJMainLayer(tf.keras.layers.Layer):
                 all_hidden_states = all_hidden_states + (tf.reshape(hidden_states, output_shape),)
 
             outputs = block(
-                hidden_states,
-                layer_past,
-                attention_mask,
-                head_mask[i],
-                use_cache,
-                output_attentions,
+                hidden_states=hidden_states,
+                layer_past=layer_past,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                head_mask=head_mask[i],
+                use_cache=use_cache,
+                output_attentions=output_attentions,
                 training=training,
             )
 
@@ -722,8 +719,6 @@ class TFGPTJForCausalLM(TFGPTJPreTrainedModel, TFCausalLanguageModelingLoss):
         self.lm_head = tf.keras.layers.Dense(
             config.vocab_size, kernel_initializer=get_initializer(config.initializer_range), name="lm_head"
         )
-        # TODO (Joao): investigate why GPTJ has numerical issues in XLA generate
-        self.supports_xla_generation = False
 
     def get_output_embeddings(self):
         return self.lm_head
@@ -731,25 +726,21 @@ class TFGPTJForCausalLM(TFGPTJPreTrainedModel, TFCausalLanguageModelingLoss):
     def set_output_embeddings(self, new_embeddings):
         self.lm_head = new_embeddings
 
-    def prepare_inputs_for_generation(self, inputs, past=None, use_cache=None, use_xla=False, **kwargs):
-        # TODO: (Joao) after the TF generator is complete, update GPT2 TF generation to match PT's. NB -- some GPT2
-        # tests will need to be fixed after the change
-
+    def prepare_inputs_for_generation(self, inputs, past=None, use_cache=None, **kwargs):
+        token_type_ids = kwargs.get("token_type_ids", None)
         # only last token for inputs_ids if past is defined in kwargs
         if past:
             inputs = tf.expand_dims(inputs[:, -1], -1)
+            if token_type_ids is not None:
+                token_type_ids = tf.expand_dims(token_type_ids[:, -1], -1)
+
+        position_ids = kwargs.get("position_ids", None)
+        attention_mask = kwargs.get("attention_mask", None)
 
-        # TODO(pvp, Joao) - this `if use_xla` statement can be removed, but is left
-        # for a future PR to not change too many things for now.
-        # All statements in this if case apply for both xla and non-xla (as they already do in PyTorch)
-        position_ids = None
-        attention_mask = None
-        if use_xla:
-            attention_mask = kwargs.get("attention_mask", None)
-            if past is not None and attention_mask is not None:
-                position_ids = tf.reduce_sum(attention_mask, axis=1, keepdims=True) - 1
-            elif attention_mask is not None:
-                position_ids = tf.math.cumsum(attention_mask, axis=1, exclusive=True)
+        if attention_mask is not None and position_ids is None:
+            position_ids = tf.math.cumsum(attention_mask, axis=-1, exclusive=True)
+            if past:
+                position_ids = tf.expand_dims(position_ids[:, -1], -1)
 
         return {
             "input_ids": inputs,
@@ -757,6 +748,7 @@ class TFGPTJForCausalLM(TFGPTJPreTrainedModel, TFCausalLanguageModelingLoss):
             "position_ids": position_ids,
             "past": past,
             "use_cache": use_cache,
+            "token_type_ids": token_type_ids,
         }
 
     @unpack_inputs

tests/models/gptj/test_modeling_tf_gptj.py

@@ -13,7 +13,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import datetime
 import unittest
 
 from transformers import AutoTokenizer, GPTJConfig, is_tf_available
@@ -48,6 +47,7 @@ class TFGPTJModelTester:
         self.use_mc_token_ids = True
         self.vocab_size = 99
         self.hidden_size = 32
+        self.rotary_dim = 4
         self.num_hidden_layers = 5
         self.num_attention_heads = 4
         self.intermediate_size = 37
@@ -103,6 +103,7 @@ class TFGPTJModelTester:
             bos_token_id=self.bos_token_id,
             eos_token_id=self.eos_token_id,
             pad_token_id=self.pad_token_id,
+            rotary_dim=self.rotary_dim,
             return_dict=True,
         )
 
@@ -359,10 +360,10 @@ class TFGPTJModelTest(TFModelTesterMixin, TFCoreModelTesterMixin, unittest.TestC
 
 
 @require_tf
+@tooslow
+# Marked as @tooslow due to GPU OOM -- but still useful to run locally. Requires ~39GB of RAM.
 class TFGPTJModelLanguageGenerationTest(unittest.TestCase):
-    @tooslow
     def test_lm_generate_gptj(self):
-        # Marked as @tooslow due to GPU OOM
         model = TFGPTJForCausalLM.from_pretrained("EleutherAI/gpt-j-6B", from_pt=True)
         input_ids = tf.convert_to_tensor([[464, 3290]], dtype=tf.int32)  # The dog
         # fmt: off
@@ -372,74 +373,20 @@ class TFGPTJModelLanguageGenerationTest(unittest.TestCase):
         output_ids = model.generate(input_ids, do_sample=False)
         self.assertListEqual(output_ids[0].numpy().tolist(), expected_output_ids)
 
-    @tooslow
     def test_gptj_sample(self):
-        # Marked as @tooslow due to GPU OOM (issue #13676)
         tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-j-6B", revision="float16")
         model = TFGPTJForCausalLM.from_pretrained("EleutherAI/gpt-j-6B", revision="float16", from_pt=True)
 
-        tf.random.set_seed(0)
-        tokenized = tokenizer("Today is a nice day and", return_tensors="tf", return_token_type_ids=True)
-        input_ids, token_type_ids = tokenized.input_ids, tokenized.token_type_ids
-        output_ids = model.generate(input_ids, do_sample=True)
+        tokenized = tokenizer("Today is a nice day and", return_tensors="tf")
+        # forces the generation to happen on CPU, to avoid GPU-related quirks
+        with tf.device(":/CPU:0"):
+            output_ids = model.generate(**tokenized, do_sample=True, seed=[42, 0])
         output_str = tokenizer.decode(output_ids[0], skip_special_tokens=True)
 
-        output_seq = model.generate(input_ids=input_ids, do_sample=True, num_return_sequences=5)
-        output_seq_tt = model.generate(
-            input_ids=input_ids, token_type_ids=token_type_ids, do_sample=True, num_return_sequences=5
-        )
-        output_seq_strs = tokenizer.batch_decode(output_seq, skip_special_tokens=True)
-        output_seq_tt_strs = tokenizer.batch_decode(output_seq_tt, skip_special_tokens=True)
-
-        EXPECTED_OUTPUT_STR = "Today is a nice day and I am taking an hour to sit in the hammock and just enjoy"
-
+        EXPECTED_OUTPUT_STR = "Today is a nice day and I’m going to go for a walk. I’"
         self.assertEqual(output_str, EXPECTED_OUTPUT_STR)
-        self.assertTrue(
-            all([output_seq_strs[idx] != output_seq_tt_strs[idx] for idx in range(len(output_seq_tt_strs))])
-        )  # token_type_ids should change output
 
-    @slow
-    @unittest.skip(reason="TF generate currently has no time-based stopping criteria")
-    def test_gptj_sample_max_time(self):
-        tokenizer = AutoTokenizer.from_pretrained("anton-l/gpt-j-tiny-random")
-        model = TFGPTJForCausalLM.from_pretrained("anton-l/gpt-j-tiny-random", from_pt=True)
-
-        input_ids = tokenizer("Today is a nice day and", return_tensors="tf", return_token_type_ids=True).input_ids
-
-        MAX_TIME = 0.5
-
-        start = datetime.datetime.now()
-        model.generate(input_ids, do_sample=True, max_time=MAX_TIME, max_length=256)
-        duration = datetime.datetime.now() - start
-        self.assertGreater(duration, datetime.timedelta(seconds=MAX_TIME))
-        self.assertLess(duration, datetime.timedelta(seconds=1.5 * MAX_TIME))
-
-        start = datetime.datetime.now()
-        model.generate(input_ids, do_sample=False, max_time=MAX_TIME, max_length=256)
-        duration = datetime.datetime.now() - start
-        self.assertGreater(duration, datetime.timedelta(seconds=MAX_TIME))
-        self.assertLess(duration, datetime.timedelta(seconds=1.5 * MAX_TIME))
-
-        start = datetime.datetime.now()
-        model.generate(input_ids, do_sample=False, num_beams=2, max_time=MAX_TIME, max_length=256)
-        duration = datetime.datetime.now() - start
-        self.assertGreater(duration, datetime.timedelta(seconds=MAX_TIME))
-        self.assertLess(duration, datetime.timedelta(seconds=1.5 * MAX_TIME))
-
-        start = datetime.datetime.now()
-        model.generate(input_ids, do_sample=True, num_beams=2, max_time=MAX_TIME, max_length=256)
-        duration = datetime.datetime.now() - start
-        self.assertGreater(duration, datetime.timedelta(seconds=MAX_TIME))
-        self.assertLess(duration, datetime.timedelta(seconds=1.5 * MAX_TIME))
-
-        start = datetime.datetime.now()
-        model.generate(input_ids, do_sample=False, max_time=None, max_length=256)
-        duration = datetime.datetime.now() - start
-        self.assertGreater(duration, datetime.timedelta(seconds=1.5 * MAX_TIME))
-
-    @tooslow
-    def test_batch_generation(self):
-        # Marked as @tooslow due to GPU OOM
+    def _get_beam_search_test_objects(self):
         model = TFGPTJForCausalLM.from_pretrained("EleutherAI/gpt-j-6B", revision="float16", from_pt=True)
         tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-j-6B", revision="float16")
 
@@ -454,42 +401,46 @@ class TFGPTJModelLanguageGenerationTest(unittest.TestCase):
             "Hello, my dog is a little",
             "Today, I",
         ]
+        expected_output_sentences = [
+            "Hello, my dog is a little over a year old and has been diagnosed with hip dysplasia",
+            "Today, I’m going to be talking about a topic that’",
+        ]
+        return model, tokenizer, sentences, expected_output_sentences
 
-        inputs = tokenizer(sentences, return_tensors="tf", padding=True)
-        input_ids = inputs["input_ids"]
-        token_type_ids = tf.concat(
-            [
-                tf.zeros((input_ids.shape[0], input_ids.shape[1] - 1), dtype=tf.int64),
-                500 * tf.ones((input_ids.shape[0], 1), dtype=tf.int64),
-            ],
-            axis=-1,
-        )
+    def test_batch_beam_search(self):
+        # Confirms that we get the expected results with left-padded beam search
+        model, tokenizer, sentences, expected_output_sentences = self._get_beam_search_test_objects()
 
-        outputs = model.generate(input_ids=input_ids, attention_mask=inputs["attention_mask"])
-        outputs_tt = model.generate(
-            input_ids=input_ids,
-            attention_mask=inputs["attention_mask"],
-            token_type_ids=token_type_ids,
-        )
+        inputs = tokenizer(sentences, return_tensors="tf", padding=True)
+        outputs = model.generate(**inputs, do_sample=False, num_beams=2)
+        batch_out_sentence = tokenizer.batch_decode(outputs, skip_special_tokens=True)
+        self.assertListEqual(expected_output_sentences, batch_out_sentence)
 
-        inputs_non_padded = tokenizer(sentences[0], return_tensors="tf").input_ids
-        output_non_padded = model.generate(input_ids=inputs_non_padded)
+    def test_batch_left_padding(self):
+        # Confirms that left-padding is working properly
+        model, tokenizer, sentences, expected_output_sentences = self._get_beam_search_test_objects()
 
+        inputs = tokenizer(sentences, return_tensors="tf", padding=True)
+        inputs_non_padded = tokenizer(sentences[0], return_tensors="tf")
+        output_non_padded = model.generate(**inputs_non_padded, do_sample=False, num_beams=2)
         num_paddings = (
-            shape_list(inputs_non_padded)[-1] - tf.reduce_sum(tf.cast(inputs["attention_mask"][-1], tf.int64)).numpy()
+            shape_list(inputs_non_padded["input_ids"])[-1]
+            - tf.reduce_sum(tf.cast(inputs["attention_mask"][-1], tf.int64)).numpy()
+        )
+        inputs_padded = tokenizer(sentences[1], return_tensors="tf")
+        output_padded = model.generate(
+            **inputs_padded, do_sample=False, num_beams=2, max_length=model.config.max_length - num_paddings
         )
-        inputs_padded = tokenizer(sentences[1], return_tensors="tf").input_ids
-        output_padded = model.generate(input_ids=inputs_padded, max_length=model.config.max_length - num_paddings)
-
-        batch_out_sentence = tokenizer.batch_decode(outputs, skip_special_tokens=True)
-        batch_out_sentence_tt = tokenizer.batch_decode(outputs_tt, skip_special_tokens=True)
         non_padded_sentence = tokenizer.decode(output_non_padded[0], skip_special_tokens=True)
         padded_sentence = tokenizer.decode(output_padded[0], skip_special_tokens=True)
+        self.assertListEqual(expected_output_sentences, [non_padded_sentence, padded_sentence])
 
-        expected_output_sentence = [
-            "Hello, my dog is a little over a year old and has been diagnosed with a heart murmur",
-            "Today, I’m going to share with you a few of my favorite",
-        ]
-        self.assertListEqual(expected_output_sentence, batch_out_sentence)
-        self.assertTrue(batch_out_sentence_tt != batch_out_sentence)  # token_type_ids should change output
-        self.assertListEqual(expected_output_sentence, [non_padded_sentence, padded_sentence])
+    def test_xla_beam_search(self):
+        # Confirms that XLA is working properly
+        model, tokenizer, sentences, expected_output_sentences = self._get_beam_search_test_objects()
+
+        inputs = tokenizer(sentences, return_tensors="tf", padding=True)
+        xla_generate = tf.function(model.generate, jit_compile=True)
+        outputs_xla = xla_generate(**inputs, do_sample=False, num_beams=2)
+        xla_sentence = tokenizer.batch_decode(outputs_xla, skip_special_tokens=True)
+        self.assertListEqual(expected_output_sentences, xla_sentence)