Use original TPU backend if possible

tpu_mtj_backend.py

@@ -60,7 +60,7 @@ def __batch_xmap(shard_dim=1):
     return inner
 
 
-def apply_repetition_penalty(logits, tokens, repetition_penalty):
+def apply_repetition_penalty_dynamic(logits, tokens, repetition_penalty):
     '''
     This gets called by generate_loop_fn to apply repetition penalty
     to the 1D array logits using the provided 1D array of tokens to penalize
@@ -85,7 +85,7 @@ def apply_repetition_penalty(logits, tokens, repetition_penalty):
     logits[tokens] = penalty_logits
     return logits
 
-def kobold_sample(key, logits, top_p=0.9, temp=0.5, top_k=0, tfs=1.0):
+def kobold_sample_dynamic(key, logits, top_p=0.9, temp=0.5, top_k=0, tfs=1.0):
     '''
     This gets called by generate_loop_fn to apply a series of 4 filters
     to the logits (top-k, then top-p, then TFS, then temperature) before
@@ -183,6 +183,127 @@ def kobold_sample(key, logits, top_p=0.9, temp=0.5, top_k=0, tfs=1.0):
     # probability distribution)
     return jax.random.categorical(key, logits, -1).astype(np.uint32)
 
+def apply_repetition_penalty_static(logits, tokens, repetition_penalty):
+    '''
+    This gets called by generate_loop_fn to apply repetition penalty
+    to the 1D array logits using the provided 1D array of tokens to penalize
+    '''
+    # Make a new array with the same length as the tokens array but with
+    # each element replaced by the value at the corresponding index in the
+    # logits array; e.g.
+    # if logits is [77, 5, 3, 98] and tokens is [0, 1, 2, 3, 2, 3, 1],
+    # then penalty_logits will be [77, 5, 3, 98, 3, 98, 5]
+    penalty_logits = jnp.take(logits, tokens)
+    # Divide positive values by repetition_penalty and multiply negative
+    # values by repetition_penalty (the academic publication that described
+    # this technique actually just only divided, but that would cause tokens
+    # with negative logits to become more likely, which is obviously wrong)
+    penalty_logits = jnp.where(
+        penalty_logits > 0,
+        penalty_logits/repetition_penalty,
+        penalty_logits*repetition_penalty,
+    )
+    # Finally, put those penalized logit values back into their original
+    # positions in the logits array
+    return logits.at[tokens].set(penalty_logits)
+
+def kobold_sample_static(key, logits, top_p=0.9, temp=0.5, top_k=0, tfs=1.0):
+    '''
+    This gets called by generate_loop_fn to apply a series of 4 filters
+    to the logits (top-k, then top-p, then TFS, then temperature) before
+    picking one token using the modified logits
+    '''
+    # Top-k (keep only the k tokens with the highest logits and remove
+    # the rest, by setting their logits to negative infinity)
+    def top_k_filter(logits):
+        # After sorting the logits array in descending order,
+        # sorted_indices_to_remove is a 1D array that is True for tokens
+        # in the sorted logits array we want to remove and False for ones
+        # we want to keep, in this case the first top_k elements will be
+        # False and the rest will be True
+        sorted_indices_to_remove = jnp.arange(len(logits)) >= top_k
+        # Unsort the logits array back to its original configuration and
+        # remove tokens we need to remove
+        _, indices_to_remove = jax.lax.sort_key_val(
+            jnp.argsort(-logits),
+            sorted_indices_to_remove,
+        )
+        return jnp.where(indices_to_remove, -jnp.inf, logits)
+    logits = jax.lax.cond(top_k > 0, top_k_filter, lambda x: x, logits)
+    # Top-p (after sorting the remaining tokens again in descending order of
+    # logit, remove the ones that have cumulative softmax probability
+    # greater than p)
+    def top_p_filter(logits):
+        # Sort the logits array in descending order, replace every element
+        # with e (Euler's number) to the power of that element, and divide
+        # each element of the new array by the sum of the elements in the
+        # new array
+        sorted_logits = -jnp.sort(-logits)
+        probabilities = jax.nn.softmax(sorted_logits)
+        # Calculate cumulative_probabilities as the prefix-sum array of
+        # probabilities
+        cumulative_probabilities = jnp.cumsum(probabilities, axis=-1)
+        # We want to remove tokens with cumulative probability higher
+        # than top_p
+        sorted_indices_to_remove = cumulative_probabilities > top_p
+        # Don't ever remove the token with the highest logit, even if
+        # the probability is higher than top_p
+        sorted_indices_to_remove = sorted_indices_to_remove.at[0].set(False)
+        # Unsort and remove
+        _, indices_to_remove = jax.lax.sort_key_val(
+            jnp.argsort(-logits),
+            sorted_indices_to_remove,
+        )
+        return jnp.where(indices_to_remove, -jnp.inf, logits)
+    logits = jax.lax.cond(top_p < 1.0, top_p_filter, lambda x: x, logits)
+    # Tail free sampling (basically top-p a second time on remaining tokens
+    # except it's the "cumulative normalized absolute second finite
+    # differences of the softmax probabilities" instead of just the
+    # cumulative softmax probabilities)
+    def tail_free_filter(logits):
+        # Sort in descending order
+        sorted_logits = -jnp.sort(-logits)
+        # Softmax again
+        probabilities = jax.nn.softmax(sorted_logits)
+        # Calculate the second finite differences of that array (i.e.
+        # calculate the difference array and then calculate the difference
+        # array of the difference array)
+        d2 = jnp.diff(jnp.diff(probabilities))
+        # Get the absolute values of all those second finite differences
+        d2 = jnp.abs(d2)
+        # Normalize (all elements in the array are divided by the sum of the
+        # array's elements)
+        d2 = d2 / d2.sum(axis=-1, keepdims=True)
+        # Get the prefix-sum array
+        cumulative_d2 = jnp.cumsum(d2, axis=-1)
+        # We will remove the tokens with a cumulative normalized absolute
+        # second finite difference larger than the TFS value
+        sorted_indices_to_remove = cumulative_d2 > tfs
+        # Don't remove the token with the highest logit
+        sorted_indices_to_remove = sorted_indices_to_remove.at[0].set(False)
+        # Since the d2 array has two fewer elements than the logits array,
+        # we'll add two extra Trues to the end
+        sorted_indices_to_remove = jnp.pad(
+            sorted_indices_to_remove,
+            (0, 2),
+            constant_values=True,
+        )
+        # Unsort and remove
+        _, indices_to_remove = jax.lax.sort_key_val(
+            jnp.argsort(-logits),
+            sorted_indices_to_remove,
+        )
+        return jnp.where(indices_to_remove, -jnp.inf, logits)
+    logits = jax.lax.cond(tfs < 1.0, tail_free_filter, lambda x: x, logits)
+    # Temperature (just divide the logits by the temperature)
+    def temp_filter(logits):
+        return logits / temp
+    logits = jax.lax.cond(True, temp_filter, lambda x: x, logits)
+    # Finally, pick one token using the softmax thingy again (it gives
+    # an array whose elements sum to 1 so it can be used nicely as a
+    # probability distribution)
+    return jax.random.categorical(key, logits, -1).astype(jnp.uint32)
+
 pad_token_id = 50256
 
 def sample_func(data, key, numseqs_aux, badwords, repetition_penalty, sampler_options):
@@ -192,11 +313,11 @@ def sample_func(data, key, numseqs_aux, badwords, repetition_penalty, sampler_op
         generated, generated_index, logits, _ = carry[0][0]
         sample_key = carry[1]
         # Get the pseudo-random number generator key that will
-        # be used by kobold_sample to randomly pick a token
+        # be used by kobold_sample_dynamic to randomly pick a token
         sample_key, new_key = jax.random.split(sample_key, num=2)
         # Apply repetition penalty to all tokens that are
         # currently inside the "generated" array
-        logits = apply_repetition_penalty(
+        logits = apply_repetition_penalty_dynamic(
             logits,
             generated,
             repetition_penalty
@@ -205,11 +326,11 @@ def sample_func(data, key, numseqs_aux, badwords, repetition_penalty, sampler_op
         # their logits to negative infinity which effectively
         # makes their probabilities of being chosen zero
         logits[badwords] = -np.inf
-        # Use the sampler (kobold_sample) to pick one token
+        # Use the sampler (kobold_sample_dynamic) to pick one token
         # based on the logits array as a 0D uint32 array
         # (higher logit means higher probability of being
         # picked, non-linearly)
-        next_token = kobold_sample(
+        next_token = kobold_sample_dynamic(
             sample_key,
             logits,
             **sampler_options,
@@ -236,6 +357,100 @@ class PenalizingCausalTransformer(CausalTransformer):
     def __init__(self, config):
         # Initialize
         super().__init__(config)
+        def generate_static(state, key, ctx, ctx_length, gen_length, numseqs_aux, sampler_options, soft_embeddings=None):
+            numseqs = numseqs_aux.shape[0]
+            # These are the tokens that we don't want the AI to ever write
+            self.badwords = jnp.array([6880, 50256, 42496, 4613, 17414, 22039, 16410, 27, 29, 38430, 37922, 15913, 24618, 28725, 58, 47175, 36937, 26700, 12878, 16471, 37981, 5218, 29795, 13412, 45160, 3693, 49778, 4211, 20598, 36475, 33409, 44167, 32406, 29847, 29342, 42669, 685, 25787, 7359, 3784, 5320, 33994, 33490, 34516, 43734, 17635, 24293, 9959, 23785, 21737, 28401, 18161, 26358, 32509, 1279, 38155, 18189, 26894, 6927, 14610, 23834, 11037, 14631, 26933, 46904, 22330, 25915, 47934, 38214, 1875, 14692, 41832, 13163, 25970, 29565, 44926, 19841, 37250, 49029, 9609, 44438, 16791, 17816, 30109, 41888, 47527, 42924, 23984, 49074, 33717, 31161, 49082, 30138, 31175, 12240, 14804, 7131, 26076, 33250, 3556, 38381, 36338, 32756, 46581, 17912, 49146])
+            @hk.transform
+            def generate_sample(context, ctx_length):
+                # Give the initial context to the transformer
+                transformer = CausalTransformerShard(config)
+                def generate_initial_scan_fn(sequence_index, _):
+                    _, initial_state = transformer.generate_initial(context, ctx_length, soft_embeddings=soft_embeddings)
+                    # The "generated" array will contain the tokens from the
+                    # context as well as the tokens picked by the sampler at
+                    # each stage, padded with a bunch of 50256s, so we know
+                    # which tokens have to be repetition penalized
+                    generated = jnp.pad(context, (0, config["seq"]), constant_values=pad_token_id)  # Let it start off with just the 2048 context tokens, plus some 50256s which will be eventually filled with sampler-chosen tokens
+                    generated_index = config["seq"]
+                    # Add that information to generate_loop_fn's starting state
+                    initial_state = (generated, generated_index, sequence_index) + initial_state
+                    return sequence_index+1, initial_state
+                _, initial_states = jax.lax.scan(generate_initial_scan_fn, 0, None, numseqs)
+                sample_key = initial_states[-1][0]
+                initial_states = list(jax.tree_map(lambda x: x[i], initial_states[:-1]) for i in range(numseqs))
+                # Get repetition penalty from the arguments
+                repetition_penalty = sampler_options.pop('repetition_penalty', None)
+                # This is the main generation loop
+                def generate_loop_fn(carry):
+                    # Unpack current generate_loop_fn state
+                    generated, generated_index, sequence_index, next_token, decode_state = carry[0][0]
+                    sample_key = carry[1]
+                    # Get the pseudo-random number generator key that will
+                    # be used by kobold_sample_static to randomly pick a token
+                    sample_key, new_key = jax.random.split(sample_key)
+                    # Give the context to the model and get the logits it
+                    # spits out
+                    # (a 2D array with 1 row and 50400 columns representing
+                    # how strongly it thinks each of the 50257 tokens in its
+                    # vocabulary should be appended to the context, followed
+                    # by 143 apparently useless columns ???)
+                    logits, new_state = transformer.generate_once(next_token, decode_state, soft_embeddings=soft_embeddings)
+                    # Verify that logits does indeed have that many rows and
+                    # columns (if you get an error here, pray for mercy)
+                    assert logits.shape == (1, config["n_vocab"])
+                    # Flatten it into a 1D array to make it easier to use
+                    logits = logits[0]
+                    # Apply repetition penalty to all tokens that are
+                    # currently inside the "generated" array
+                    if repetition_penalty is not None:
+                        logits = apply_repetition_penalty_static(
+                            logits,
+                            generated,
+                            repetition_penalty
+                        )
+                    # Remove any tokens in the badwords list by setting
+                    # their logits to negative infinity which effectively
+                    # makes their probabilities of being chosen zero
+                    logits = logits.at[self.badwords].set(-jnp.inf)
+                    # Use the sampler (kobold_sample_static) to pick one token
+                    # based on the logits array as a 0D uint32 array
+                    # (higher logit means higher probability of being
+                    # picked, non-linearly)
+                    next_token = kobold_sample_static(
+                        sample_key,
+                        logits,
+                        **sampler_options,
+                    )
+                    # Remember what token was picked
+                    generated = generated.at[generated_index].set(next_token)
+                    generated_index += 1
+                    # Re-pack the current generate_loop_fn's state so we can
+                    # get back the same variables the next time
+                    carry[0][0] = (generated, generated_index, sequence_index, next_token[jnp.newaxis], new_state)
+                    carry[0].append(carry[0].pop(0))
+                    return carry[0], new_key
+                return jax.lax.while_loop(
+                    lambda carry: carry[0][0][1] - config["seq"] < gen_length,
+                    generate_loop_fn,
+                    (initial_states, sample_key),
+                )
+            return generate_sample.apply(state["params"], key, ctx, ctx_length)
+        self.generate_static_xmap = jax.experimental.maps.xmap(
+            fun=generate_static,
+            in_axes=(
+                ["shard", ...],
+                ["batch", ...],
+                ["batch", ...],
+                ["batch", ...],
+                ["batch", ...],
+                ["batch", ...],
+                ["batch", ...],
+                ["shard", ...],
+            ),
+            out_axes=["shard", "batch", ...],
+            axis_resources={'shard': 'mp', 'batch': 'dp'},
+        )
         def generate_initial(state, key, ctx, ctx_length, numseqs_aux, soft_embeddings=None):
             numseqs = numseqs_aux.shape[0]
             @hk.transform
@@ -314,7 +529,7 @@ class PenalizingCausalTransformer(CausalTransformer):
             out_axes=["shard", "batch", ...],
             axis_resources={'shard': 'mp', 'batch': 'dp'},
         )
-    def generate(self, ctx, ctx_length, gen_length, numseqs, sampler_options, return_logits=False, soft_embeddings=None, excluded_world_info=None, use_callback=True):
+    def generate_dynamic(self, ctx, ctx_length, gen_length, numseqs, sampler_options, return_logits=False, soft_embeddings=None, excluded_world_info=None, use_callback=True):
         assert excluded_world_info is not None
         assert not return_logits
         assert gen_length.ndim == 1
@@ -360,9 +575,24 @@ class PenalizingCausalTransformer(CausalTransformer):
             else:
                 break
         return sample_data, n_generated, regeneration_required, halt
+    def generate_static(self, ctx, ctx_length, gen_length, numseqs, sampler_options, return_logits=False, soft_embeddings=None):
+        assert not return_logits
+        key = hk.PRNGSequence(random.randint(0, 2 ** 60))
+        batch_size = ctx.shape[0]
+        self.batch_size = batch_size
+        return self.generate_static_xmap(
+            self.state,
+            jnp.array(key.take(batch_size)),
+            ctx,
+            np.array(ctx_length, dtype=np.uint32),
+            np.array(gen_length, dtype=np.uint32),
+            np.empty((batch_size, numseqs), dtype=np.uint8),
+            sampler_options,
+            soft_embeddings,
+        )
 
 
-def infer(
+def infer_dynamic(
     context: np.array,
     top_p=0.9,
     temp=0.5,
@@ -394,7 +624,7 @@ def infer(
         "repetition_penalty": float(repetition_penalty),
         "top_k": int(top_k),
     }
-    output = network.generate(
+    output = network.generate_dynamic(
         batched_tokens,
         np.ones(total_batch, dtype=np.uint32) * provided_ctx,
         np.ones(total_batch, dtype=np.uint32) * gen_len,
@@ -408,6 +638,47 @@ def infer(
         samples.append(out[0][params["seq"] : params["seq"] + gen_len])
     return (samples,) + output[1:]
 
+def infer_static(
+    context: np.array,
+    top_p=0.9,
+    temp=0.5,
+    top_k=0,
+    tfs=1.0,
+    repetition_penalty=1.0,
+    numseqs=1,
+    gen_len=80,
+    soft_embeddings: Optional[np.array] = None,
+    soft_tokens: Optional[np.array] = None,
+) -> List[np.array]:
+    maps.thread_resources.env = thread_resources_env
+    total_batch = 1
+    tokens = context
+    if(soft_tokens is not None):
+        tokens = np.uint32(np.concatenate((soft_tokens, tokens)))
+    provided_ctx = tokens.shape[0]
+    pad_amount = seq - provided_ctx
+    padded_tokens = np.pad(tokens, ((pad_amount, 0),), constant_values=pad_token_id)
+    batched_tokens = np.array([padded_tokens] * total_batch)
+    samples = []
+    batched_generator_params = {
+        "temp": temp * np.ones(total_batch),
+        "top_p": top_p * np.ones(total_batch),
+        "tfs": tfs * np.ones(total_batch),
+        "repetition_penalty": repetition_penalty * np.ones(total_batch),
+        "top_k": np.full(total_batch, top_k, dtype=np.uint32)
+    }
+    output = network.generate_static(
+        batched_tokens,
+        np.ones(total_batch, dtype=np.uint32) * provided_ctx,
+        np.ones(total_batch, dtype=np.uint32) * gen_len,
+        numseqs,
+        batched_generator_params,
+        soft_embeddings=soft_embeddings,
+    )[0]
+    for o in output:
+        samples.append(o[0][0, 0, params["seq"] : params["seq"] + gen_len])
+    return samples
+
 
 def load_model(path: str, driver_version="tpu_driver0.1_dev20210607", **kwargs) -> None:
     global thread_resources_env, seq, tokenizer, network, params