Upload ProGen2 base BF16 MLX conversion

README.md

@@ -0,0 +1,51 @@
+---
+license: bsd-3-clause
+library_name: mlx
+pipeline_tag: text-generation
+base_model: hugohrban/progen2-base
+tags:
+- progen
+- progen2
+- protein-language-model
+- mlx
+- mlx-lm
+- safetensors
+- bfloat16
+- icl-many-replication
+---
+
+# ProGen2 Base BF16 MLX Conversion
+
+This repository contains a BF16 MLX-LM conversion of
+[`hugohrban/progen2-base`](https://huggingface.co/hugohrban/progen2-base), a
+mirror of the base ProGen2 model by Nijkamp et al.
+
+The source checkpoint was downloaded from Hugging Face revision
+`71228cbfca5960f9fab5775f378bba3673af9f00` and converted from FP32 to BF16
+safetensors. The converted tensor file contains 764,803,616 BF16 parameters.
+
+This conversion was prepared for the standalone replication artifact for
+*Many Next Token Predictors are In-context Learners*.
+
+## Files
+
+- `model.safetensors`: BF16 converted ProGen2-base weights.
+- `progen2_mlx.py`: MLX-LM custom model implementation.
+- `config.json`, tokenizer files, and ProGen custom-code files copied from the
+  source model with the MLX model-file entry added.
+
+## Loading with MLX-LM
+
+```python
+from mlx_lm import load
+
+model, tokenizer = load("N8Programs/ProGen2-base-bf16")
+```
+
+## Original Model
+
+Please cite and follow the terms of the upstream ProGen2 work and the source
+model repository:
+
+- Source model: https://huggingface.co/hugohrban/progen2-base
+- Paper: https://arxiv.org/abs/2206.13517

config.json

@@ -0,0 +1,49 @@
+{
+  "_name_or_path": "hugohrban/progen2-base",
+  "activation_function": "gelu_new",
+  "architectures": [
+    "ProGenForCausalLM"
+  ],
+  "attn_pdrop": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_progen.ProGenConfig",
+    "AutoModelForCausalLM": "modeling_progen.ProGenForCausalLM"
+  },
+  "bos_token_id": 1,
+  "embd_pdrop": 0.0,
+  "embed_dim": 1536,
+  "eos_token_id": 2,
+  "gradient_checkpointing": false,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "progen",
+  "n_head": 16,
+  "n_inner": null,
+  "n_layer": 27,
+  "n_positions": 2048,
+  "resid_pdrop": 0.0,
+  "rotary_dim": 48,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50,
+      "temperature": 1.0
+    }
+  },
+  "tokenizer_class": "GPT2Tokenizer",
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.40.0",
+  "use_cache": true,
+  "vocab_size": 32,
+  "vocab_size_emb": 32,
+  "vocab_size_lm_head": 32,
+  "model_file": "progen2_mlx.py",
+  "dtype": "bfloat16",
+  "pad_token_id": 0
+}

configuration_progen.py

@@ -0,0 +1,72 @@
+# coding=utf-8
+# Copyright 2021 The EleutherAI and HuggingFace Teams. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Modified configuration implementation based on https://github.com/huggingface/transformers/blob/main/src/transformers/models/gptj/configuration_gptj.py
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class ProGenConfig(PretrainedConfig):
+    model_type = "progen"
+
+    def __init__(
+        self,
+        vocab_size_emb=32,
+        vocab_size_lm_head=32,
+        n_positions=1024,
+        embed_dim=1024,
+        n_layer=12,
+        n_head=16,
+        rotary_dim=32,
+        n_inner=None,
+        activation_function="gelu_new",
+        resid_pdrop=0.0,
+        embd_pdrop=0.0,
+        attn_pdrop=0.0,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        scale_attn_weights=True,
+        gradient_checkpointing=False,
+        use_cache=True,
+        bos_token_id=1,
+        eos_token_id=2,
+        **kwargs
+    ):
+        super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size_emb = vocab_size_emb                # input vocab size
+        self.vocab_size_lm_head = vocab_size_lm_head        # output vocab size
+        self.n_positions = n_positions                      # context window size
+        self.embed_dim = embed_dim
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_inner = n_inner                              # inner dimension of the MLP
+        self.rotary_dim = rotary_dim
+        self.activation_function = activation_function
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.gradient_checkpointing = gradient_checkpointing
+        self.scale_attn_weights = scale_attn_weights
+        self.use_cache = use_cache
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "transformers_version": "4.40.0"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc2ee1fe37ec29e2371b37ba9f06ab468cce7e30d6d8b3eb0b99a9fee8d23ccb
+size 1529630800

modeling_progen.py

@@ -0,0 +1,654 @@
+# coding=utf-8
+# Copyright 2021 The EleutherAI and HuggingFace Teams. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License atí
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Modified forward-pass implementation based on https://github.com/huggingface/transformers/blob/main/src/transformers/models/gptj/modeling_gptj.py
+
+from typing import Tuple
+
+import numpy as np
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+import torch.nn.functional as F
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from .configuration_progen import ProGenConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+def fixed_pos_embedding(x, seq_dim=1, seq_len=None):
+    dim = x.shape[-1]
+    if seq_len is None:
+        seq_len = x.shape[seq_dim]
+    inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2) / dim))
+    sinusoid_inp = (
+        torch.einsum("i , j -> i j", torch.arange(seq_len), inv_freq)
+        .to(x.device)
+        .float()
+    )
+    return torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)
+
+
+def rotate_every_two(x: torch.Tensor):
+    x1 = x[:, :, :, ::2]
+    x2 = x[:, :, :, 1::2]
+    x = torch.stack((-x2, x1), axis=-1)
+    return x.flatten(-2)
+
+def apply_rotary_pos_emb(x, sincos, offset=0):
+    sin, cos = map(
+        lambda t: t[None, offset : x.shape[1] + offset, None, :].repeat_interleave(
+            2, 3
+        ),
+        sincos,
+    )
+    # einsum notation for lambda t: repeat(t[offset:x.shape[1]+offset,:], "n d -> () n () (d j)", j=2)
+    return (x * cos) + (rotate_every_two(x) * sin)
+
+
+class ProGenAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        max_positions = config.n_positions
+        self.register_buffer(
+            "bias",
+            torch.tril(
+                torch.ones((max_positions, max_positions), dtype=torch.bool)
+            ).view(1, 1, max_positions, max_positions),
+            persistent=False
+        )
+        self.register_buffer("masked_bias", torch.tensor(-1e9), persistent=False)     # approx. -inf
+
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+        self.resid_dropout = nn.Dropout(config.resid_pdrop)
+
+        self.embed_dim = config.embed_dim
+        self.num_attention_heads = config.n_head
+        self.head_dim = self.embed_dim // self.num_attention_heads
+        if self.head_dim * self.num_attention_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and `num_attention_heads`: {self.num_attention_heads})."
+            )
+        self.scale_attn = torch.sqrt(
+            torch.tensor(self.head_dim, dtype=torch.float32)
+        ).to(torch.get_default_dtype())
+        self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)
+
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
+        self.rotary_dim = None
+        if config.rotary_dim is not None:
+            self.rotary_dim = config.rotary_dim
+
+    def _split_heads(self, x: torch.Tensor, n_head, dim_head) -> torch.Tensor:
+        x = x.reshape(x.shape[:-2] + (-1,))                             # (B, T, 8 * E // 8)
+        x = x.reshape(x.shape[:-1] + (n_head, dim_head))             # (B, T, n_heads, dim_head)
+        return x
+
+    def _merge_heads(self, tensor, num_attention_heads, attn_head_size) -> torch.Tensor:
+        """
+        Merges attn_head_size dim and num_attn_heads dim into n_positions
+        """
+        if len(tensor.shape) == 5:
+            tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()
+        elif len(tensor.shape) == 4:
+            tensor = tensor.permute(0, 2, 1, 3).contiguous()
+        else:
+            raise ValueError(
+                f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}"
+            )
+        new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)
+        return tensor.view(new_shape)
+
+    def _attn(
+        self,
+        query,
+        key,
+        value,
+        attention_mask=None,
+        head_mask=None,
+    ):
+        # compute causal mask from causal mask buffer
+        query_length, key_length = query.size(-2), key.size(-2)
+        causal_mask = self.bias[
+            :, :, key_length - query_length : key_length, :key_length
+        ]
+
+        # Keep the attention weights computation in fp32 to avoid overflow issues
+        query = query.to(torch.float32)
+        key = key.to(torch.float32)
+
+        attn_weights = query @ key.transpose(-1, -2)                 # (B, n_heads, T, T)
+
+        attn_weights = attn_weights / self.scale_attn
+
+        # attend only to previous positions
+        attn_weights = torch.where(
+            causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype)
+        )
+
+        if attention_mask is not None:
+            attn_weights = attn_weights + attention_mask
+
+        attn_weights = F.softmax(attn_weights, dim=-1)
+        attn_weights = attn_weights.to(value.dtype)
+        attn_weights = self.attn_dropout(attn_weights)
+
+        if head_mask is not None:
+            attn_weights = attn_weights * head_mask
+
+        attn_output = attn_weights @ value                # (B, n_heads, T, dim_head)
+
+        return attn_output, attn_weights
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        layer_past=None,
+        head_mask=None,
+        use_cache=False,
+        output_attentions=False,
+    ):
+        qkv = self.qkv_proj(hidden_states)                                         # (B, T, 3 * E)
+
+        mp_num = 8
+        qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))                     # (B, T, 8, 3 * E // 8)
+
+        query, value, key = torch.split(qkv_split, self.embed_dim // mp_num, dim=-1) # 3 * (B, T, 8, E // 8)
+
+        query = self._split_heads(query, self.num_attention_heads, self.head_dim)  # (B, T, n_heads, dim_head)
+        key = self._split_heads(key, self.num_attention_heads, self.head_dim)      # (B, T, n_heads, dim_head)
+        value = self._split_heads(value, self.num_attention_heads, self.head_dim)  # (B, T, n_heads, dim_head)
+        value = value.permute(0, 2, 1, 3)
+
+        seq_len = key.shape[1]
+        offset = 0
+
+        if layer_past is not None:
+            offset = layer_past[0].shape[-2]
+            seq_len += offset
+
+        if self.rotary_dim is not None:
+            k_rot = key[:, :, :, : self.rotary_dim]
+            k_pass = key[:, :, :, self.rotary_dim :]
+
+            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)
+
+            key = torch.cat([k_rot, k_pass], dim=-1)
+            query = torch.cat([q_rot, q_pass], dim=-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 = key.permute(0, 2, 1, 3)
+        query = query.permute(0, 2, 1, 3)
+
+        if layer_past is not None:
+            past_key = layer_past[0]
+            past_value = layer_past[1]
+            key = torch.cat((past_key, key), dim=-2)
+            value = torch.cat((past_value, value), dim=-2)
+
+        if use_cache is True:
+            present = (key, value)
+        else:
+            present = None
+
+        # compute self-attention: softmax((Q @ K.T) / sqrt(dim_head)) @ V
+        attn_output, attn_weights = self._attn(
+            query, key, value, attention_mask, head_mask
+        )
+
+        attn_output = self._merge_heads(                               # (B, T, E)  
+            attn_output, self.num_attention_heads, self.head_dim
+        )
+
+        attn_output = self.out_proj(attn_output)
+        attn_output = self.resid_dropout(attn_output)
+
+        outputs = (attn_output, present)
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs  # a, present, (attentions)
+
+
+class ProGenMLP(nn.Module):
+    def __init__(
+        self, intermediate_size, config
+    ):  # in MLP: intermediate_size= 4 * embed_dim
+        super().__init__()
+        embed_dim = config.embed_dim
+
+        self.fc_in = nn.Linear(embed_dim, intermediate_size)
+        self.fc_out = nn.Linear(intermediate_size, embed_dim)
+
+        self.act = ACT2FN[config.activation_function]
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(self, hidden_states):
+        hidden_states = self.fc_in(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.fc_out(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        return hidden_states
+
+
+class ProGenBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * config.embed_dim
+        self.ln_1 = nn.LayerNorm(config.embed_dim, eps=config.layer_norm_epsilon)
+        self.attn = ProGenAttention(config)
+        self.mlp = ProGenMLP(inner_dim, config)
+
+    def forward(
+        self,
+        hidden_states,
+        layer_past=None,
+        attention_mask=None,
+        head_mask=None,
+        use_cache=False,
+        output_attentions=False,
+    ):
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_outputs = self.attn(
+            hidden_states,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            head_mask=head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
+        outputs = attn_outputs[1:]
+
+        feed_forward_hidden_states = self.mlp(hidden_states)            # (B, T, E)
+        hidden_states = attn_output + feed_forward_hidden_states + residual
+
+        if use_cache:
+            outputs = (hidden_states,) + outputs
+        else:
+            outputs = (hidden_states,) + outputs[1:]
+
+        return outputs  # hidden_states, present, (attentions)
+
+
+class ProGenPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = ProGenConfig
+    base_model_prefix = "transformer"
+    is_parallelizable = False
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, (nn.Linear,)):
+            # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+class ProGenModel(ProGenPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.vocab_size_emb = config.vocab_size_emb
+        self.embed_dim = config.embed_dim
+        self.wte = nn.Embedding(config.vocab_size_emb, self.embed_dim)
+        self.drop = nn.Dropout(config.embd_pdrop)
+        self.h = nn.ModuleList([ProGenBlock(config) for _ in range(config.n_layer)])
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+        self.rotary_dim = min(
+            config.rotary_dim, config.n_positions // config.n_head
+        )
+        self.init_weights()
+
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both input_ids and inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size = inputs_embeds.shape[0]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.view(-1, input_shape[-1])
+
+        if position_ids is not None:
+            position_ids = position_ids.view(-1, input_shape[-1])
+
+        if past_key_values is None:
+            past_length = 0
+            past_key_values = tuple([None] * len(self.h))
+        else:
+            past_length = past_key_values[0][0].size(-2)
+
+        if position_ids is None:
+            position_ids = torch.arange(
+                past_length,
+                input_shape[-1] + past_length,
+                dtype=torch.long,
+                device=device,
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
+
+        # Attention mask.
+        if attention_mask is not None:
+            assert batch_size > 0, "batch_size has to be defined and > 0"
+            attention_mask = attention_mask.view(batch_size, -1)
+            # We create a 3D attention mask from a 2D tensor mask.
+            # Sizes are [batch_size, 1, 1, to_seq_length]
+            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+            # this attention mask is more simple than the triangular masking of causal attention
+            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+            attention_mask = attention_mask[:, None, None, :]
+
+            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+            # masked positions, this operation will create a tensor which is 0.0 for
+            # positions we want to attend and -10000.0 for masked positions.
+            # Since we are adding it to the raw scores before the softmax, this is
+            # effectively the same as removing these entirely.
+            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+            attention_mask = (1.0 - attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x num_attention_heads x N x N
+        # head_mask has shape n_layer x batch x num_attention_heads x N x N
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.wte(input_ids)
+
+        hidden_states = inputs_embeds
+
+        if token_type_ids is not None:
+            token_type_embeds = self.wte(token_type_ids)
+            hidden_states = hidden_states + token_type_embeds
+
+        hidden_states = self.drop(hidden_states)
+
+        output_shape = input_shape + (hidden_states.size(-1),)
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if getattr(self.config, "gradient_checkpointing", False) and self.training:
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with `config.gradient_checkpointing=True`. Setting "
+                        "`use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache, output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    None,
+                    attention_mask,
+                    head_mask[i],
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=attention_mask,
+                    head_mask=head_mask[i],
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (
+                    outputs[2 if use_cache else 1],
+                )
+
+        hidden_states = self.ln_f(hidden_states)
+
+        hidden_states = hidden_states.view(*output_shape)
+        # Add last hidden state
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    presents,
+                    all_hidden_states,
+                    all_self_attentions,
+                ]
+                if v is not None
+            )
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class ProGenForCausalLM(ProGenPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [
+        r"h\.\d+\.attn\.masked_bias",
+        r"h\.\d+\.attn\.bias",
+        r"lm_head\.weight",
+    ]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = ProGenModel(config)
+        self.lm_head = nn.Linear(config.embed_dim, config.vocab_size_lm_head)
+        self.init_weights()
+
+    def prepare_inputs_for_generation(self, input_ids, past=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:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "position_ids": position_ids,
+            "attention_mask": attention_mask,
+            "token_type_ids": token_type_ids,
+        }
+
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            ``labels = input_ids`` Indices are selected in ``[-100, 0, ..., config.vocab_size]`` All labels set to
+            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``
+        """
+        return_dict = (
+            return_dict if return_dict is not None else self.config.use_return_dict
+        )
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+
+        # make sure sampling in fp16 works correctly and
+        # compute loss in fp32 to match with mesh-tf version
+        # https://github.com/EleutherAI/gpt-neo/blob/89ce74164da2fb16179106f54e2269b5da8db333/models/gpt2/gpt2.py#L179
+        lm_logits = self.lm_head(hidden_states).to(torch.float32)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
+            )
+            loss = loss.to(hidden_states.dtype)
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+        past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
+    ) -> Tuple[Tuple[torch.Tensor]]:
+        """
+        This function is used to re-order the :obj:`past_key_values` cache if
+        :meth:`~transformers.PretrainedModel.beam_search` or :meth:`~transformers.PretrainedModel.beam_sample` is
+        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(
+                past_state.index_select(0, beam_idx.to(past_state.device))
+                for past_state in layer_past
+            )
+            for layer_past in past
+        )

progen2_mlx.py

@@ -0,0 +1,254 @@
+# Copyright (c) 2026
+"""MLX-LM architecture for ProGen2 causal protein LMs."""
+
+from dataclasses import dataclass
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from mlx_lm.models.base import (
+    BaseModelArgs,
+    create_attention_mask,
+    scaled_dot_product_attention,
+)
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    vocab_size_emb: int
+    vocab_size_lm_head: int
+    n_positions: int
+    embed_dim: int
+    n_layer: int
+    n_head: int
+    rotary_dim: int = 64
+    n_inner: Optional[int] = None
+    activation_function: str = "gelu_new"
+    layer_norm_epsilon: float = 1e-5
+    bos_token_id: int = 1
+    eos_token_id: int = 2
+    pad_token_id: int = 0
+
+
+def gelu_new(x: mx.array) -> mx.array:
+    return 0.5 * x * (
+        1.0
+        + mx.tanh(0.7978845608028654 * (x + 0.044715 * mx.power(x, 3)))
+    )
+
+
+def rotate_every_two(x: mx.array) -> mx.array:
+    x1 = x[..., ::2]
+    x2 = x[..., 1::2]
+    stacked = mx.stack((-x2, x1), axis=-1)
+    return stacked.reshape(*x.shape)
+
+
+class PartialRotaryEmbedding(nn.Module):
+    def __init__(self, rotary_dim: int):
+        super().__init__()
+        self.rotary_dim = rotary_dim
+        inv_freq = 1.0 / (
+            10000
+            ** (mx.arange(0, rotary_dim, 2, dtype=mx.float32) / rotary_dim)
+        )
+        self.inv_freq = inv_freq
+
+    def __call__(self, x: mx.array, offset: int | mx.array = 0) -> mx.array:
+        seq_len = x.shape[-2]
+        offset = mx.array(offset, dtype=mx.float32)
+        positions = mx.arange(seq_len, dtype=mx.float32)
+
+        if offset.ndim == 0:
+            positions = positions + offset
+            freqs = positions[:, None] * self.inv_freq[None, :]
+            emb = mx.repeat(freqs, 2, axis=-1)
+            cos = mx.cos(emb).astype(x.dtype).reshape(
+                1,
+                1,
+                seq_len,
+                self.rotary_dim,
+            )
+            sin = mx.sin(emb).astype(x.dtype).reshape(
+                1,
+                1,
+                seq_len,
+                self.rotary_dim,
+            )
+        else:
+            positions = positions[None, :] + offset[:, None]
+            freqs = positions[:, :, None] * self.inv_freq[None, None, :]
+            emb = mx.repeat(freqs, 2, axis=-1)
+            cos = mx.cos(emb).astype(x.dtype).reshape(
+                x.shape[0],
+                1,
+                seq_len,
+                self.rotary_dim,
+            )
+            sin = mx.sin(emb).astype(x.dtype).reshape(
+                x.shape[0],
+                1,
+                seq_len,
+                self.rotary_dim,
+            )
+
+        return (x * cos) + (rotate_every_two(x) * sin)
+
+
+class ProGenAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        if args.embed_dim % args.n_head != 0:
+            raise ValueError("embed_dim must be divisible by n_head")
+        self.embed_dim = args.embed_dim
+        self.num_heads = args.n_head
+        self.head_dim = args.embed_dim // args.n_head
+        self.mp_num = 8
+        self.mp_part = args.embed_dim // self.mp_num
+        self.scale = self.head_dim**-0.5
+        self.rotary_dim = args.rotary_dim
+
+        self.qkv_proj = nn.Linear(args.embed_dim, args.embed_dim * 3, bias=False)
+        self.out_proj = nn.Linear(args.embed_dim, args.embed_dim, bias=False)
+        self.rotary = PartialRotaryEmbedding(self.rotary_dim)
+
+    def _split_heads_from_mp(self, x: mx.array) -> mx.array:
+        batch_size, seq_len = x.shape[:2]
+        x = x.reshape(batch_size, seq_len, self.embed_dim)
+        return x.reshape(
+            batch_size,
+            seq_len,
+            self.num_heads,
+            self.head_dim,
+        ).transpose(0, 2, 1, 3)
+
+    def _apply_partial_rotary(self, x: mx.array, offset: int | mx.array) -> mx.array:
+        x_rot = x[..., : self.rotary_dim]
+        x_pass = x[..., self.rotary_dim :]
+        return mx.concatenate([self.rotary(x_rot, offset=offset), x_pass], axis=-1)
+
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        mask: Optional[Any] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        batch_size, seq_len, _ = hidden_states.shape
+        qkv = self.qkv_proj(hidden_states)
+        qkv = qkv.reshape(batch_size, seq_len, self.mp_num, -1)
+
+        query, value, key = mx.split(qkv, 3, axis=-1)
+        query = self._split_heads_from_mp(query)
+        key = self._split_heads_from_mp(key)
+        value = self._split_heads_from_mp(value)
+
+        offset = 0 if cache is None else cache.offset
+        query = self._apply_partial_rotary(query, offset=offset)
+        key = self._apply_partial_rotary(key, offset=offset)
+
+        if cache is not None:
+            key, value = cache.update_and_fetch(key, value)
+
+        attn_output = scaled_dot_product_attention(
+            query,
+            key,
+            value,
+            cache=cache,
+            scale=self.scale,
+            mask=mask,
+        )
+        attn_output = attn_output.transpose(0, 2, 1, 3).reshape(
+            batch_size,
+            seq_len,
+            self.embed_dim,
+        )
+        return self.out_proj(attn_output)
+
+
+class ProGenMLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        inner_dim = args.n_inner if args.n_inner is not None else 4 * args.embed_dim
+        self.fc_in = nn.Linear(args.embed_dim, inner_dim, bias=True)
+        self.fc_out = nn.Linear(inner_dim, args.embed_dim, bias=True)
+
+    def __call__(self, hidden_states: mx.array) -> mx.array:
+        return self.fc_out(gelu_new(self.fc_in(hidden_states)))
+
+
+class ProGenBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(
+            args.embed_dim,
+            eps=args.layer_norm_epsilon,
+            affine=True,
+            bias=True,
+        )
+        self.attn = ProGenAttention(args)
+        self.mlp = ProGenMLP(args)
+
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        mask: Optional[Any] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        residual = hidden_states
+        normed = self.ln_1(hidden_states)
+        attn_output = self.attn(normed, mask=mask, cache=cache)
+        mlp_output = self.mlp(normed)
+        return residual + attn_output + mlp_output
+
+
+class ProGenModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.wte = nn.Embedding(args.vocab_size_emb, args.embed_dim)
+        self.h = [ProGenBlock(args) for _ in range(args.n_layer)]
+        self.ln_f = nn.LayerNorm(
+            args.embed_dim,
+            eps=args.layer_norm_epsilon,
+            affine=True,
+            bias=True,
+        )
+
+    def __call__(self, inputs: mx.array, cache=None) -> mx.array:
+        hidden_states = self.wte(inputs)
+        if cache is None:
+            cache = [None] * len(self.h)
+        mask = create_attention_mask(hidden_states, cache[0])
+        for block, c in zip(self.h, cache):
+            hidden_states = block(hidden_states, mask=mask, cache=c)
+        return self.ln_f(hidden_states)
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.transformer = ProGenModel(args)
+        self.lm_head = nn.Linear(args.embed_dim, args.vocab_size_lm_head, bias=True)
+
+    def __call__(self, inputs: mx.array, cache=None) -> mx.array:
+        return self.lm_head(self.transformer(inputs, cache=cache))
+
+    def sanitize(self, weights):
+        weights = dict(weights)
+        inv_freq = 1.0 / (
+            10000
+            ** (
+                mx.arange(0, self.args.rotary_dim, 2, dtype=mx.float32)
+                / self.args.rotary_dim
+            )
+        )
+        for layer_idx in range(self.args.n_layer):
+            weights[f"transformer.h.{layer_idx}.attn.rotary.inv_freq"] = inv_freq
+        return weights
+
+    @property
+    def layers(self):
+        return self.transformer.h

special_tokens_map.json

@@ -0,0 +1,6 @@
+{
+  "bos_token": "<|bos|>",
+  "eos_token": "<|eos|>",
+  "pad_token": "<|pad|>",
+  "unk_token": "<|eos|>"
+}

tokenizer.json

@@ -0,0 +1,91 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding":  null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "special": true,
+      "content": "<|pad|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false
+    },
+    {
+      "id": 1,
+      "special": true,
+      "content": "<|bos|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false
+    },
+    {
+      "id": 2,
+      "special": true,
+      "content": "<|eos|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false
+    }
+  ],
+  "normalizer": null,
+  "pre_tokenizer": {
+    "type": "ByteLevel",
+    "add_prefix_space": false,
+    "trim_offsets": true
+  },
+  "post_processor": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": true
+  },
+  "decoder": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": true
+  },
+  "model": {
+    "type": "BPE",
+    "dropout": null,
+    "unk_token": null,
+    "continuing_subword_prefix": null,
+    "end_of_word_suffix": null,
+    "fuse_unk": false,
+    "vocab": {
+      "<|pad|>": 0,
+      "<|bos|>": 1,
+      "<|eos|>": 2,
+      "1": 3,
+      "2": 4,
+      "A": 5,
+      "B": 6,
+      "C": 7,
+      "D": 8,
+      "E": 9,
+      "F": 10,
+      "G": 11,
+      "H": 12,
+      "I": 13,
+      "K": 14,
+      "L": 15,
+      "M": 16,
+      "N": 17,
+      "O": 18,
+      "P": 19,
+      "Q": 20,
+      "R": 21,
+      "S": 22,
+      "T": 23,
+      "U": 24,
+      "V": 25,
+      "W": 26,
+      "X": 27,
+      "Y": 28,
+      "Z": 29
+    },
+    "merges": []
+  }
+}

tokenizer_config.json

@@ -0,0 +1,8 @@
+{
+  "bos_token": "<|bos|>",
+  "eos_token": "<|eos|>",
+  "pad_token": "<|pad|>",
+  "unk_token": "<|eos|>",
+  "model_max_length": 2048,
+  "add_prefix_space": false
+}