Initial release of ParallelMLP adapter

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

config.json

@@ -0,0 +1,22 @@
+{
+  "architectures": [
+    "UnifiedParallelMLPForCausalLM"
+  ],
+  "base_model_name_or_path": "Qwen/Qwen2.5-1.5B-Instruct",
+  "dtype": "bfloat16",
+  "eos_token_id": 151645,
+  "hidden_size": 1536,
+  "mlp_intermediate_size": 12288,
+  "mlp_positions": [
+    24,
+    25,
+    26,
+    27,
+    28
+  ],
+  "model_type": "parallel_mlp_causal_lm",
+  "num_hidden_layers": 28,
+  "pad_token_id": 151654,
+  "rms_norm_eps": 1e-06,
+  "transformers_version": "4.57.6"
+}

configuration_parallel_mlp.py

@@ -0,0 +1,41 @@
+from transformers import PretrainedConfig, AutoConfig
+
+
+class ParallelMLPConfig(PretrainedConfig):
+    model_type = "parallel_mlp_causal_lm"
+
+    def __init__(
+        self,
+        base_model_name_or_path: str = "meta-llama/Meta-Llama-3.1-8B-Instruct",
+        mlp_positions: list[int] | None = None,
+        mlp_intermediate_size: int | None = None,
+        **kwargs,
+    ):
+        self.base_model_name_or_path = base_model_name_or_path
+        self.mlp_positions = mlp_positions if mlp_positions is not None else []
+        try:
+            base_cfg = AutoConfig.from_pretrained(base_model_name_or_path, trust_remote_code=True)
+            self.hidden_size       = getattr(base_cfg, "hidden_size", 4096)
+            self.rms_norm_eps      = getattr(base_cfg, "rms_norm_eps", 1e-5)
+            self.num_hidden_layers = getattr(base_cfg, "num_hidden_layers", 32)
+        except Exception:
+            self.hidden_size       = 4096
+            self.rms_norm_eps      = 1e-5
+            self.num_hidden_layers = 32
+
+        # Intermediate size: ~8 × hidden_size when matched to CALYREX late8th budget
+        if mlp_intermediate_size is not None:
+            self.mlp_intermediate_size = mlp_intermediate_size
+        else:
+            self.mlp_intermediate_size = 8 * self.hidden_size
+
+        for pos in self.mlp_positions:
+            if pos < 1 or pos > self.num_hidden_layers:
+                raise ValueError(
+                    f"mlp_position {pos} is out of bounds (1 to {self.num_hidden_layers})"
+                )
+
+        super().__init__(**kwargs)
+
+
+AutoConfig.register("parallel_mlp_causal_lm", ParallelMLPConfig)

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c5b5f10272685d5a2a5ff70b5f0c7d779fb63c18208c360033839e42a96b021
+size 3464975736

modeling_parallel_mlp.py

@@ -0,0 +1,86 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PreTrainedModel, AutoModelForCausalLM
+from .configuration_parallel_mlp import ParallelMLPConfig
+
+
+class ParallelMLPAdapter(nn.Module):
+    """Simple bottleneck MLP (norm → up → SiLU → down) parallel to a transformer block.
+
+    down_proj is zero-initialized so the adapter is a strict pass-through at the
+    start of training, avoiding representation shock.
+    """
+
+    def __init__(self, hidden_size: int, intermediate_size: int, rms_norm_eps: float = 1e-5):
+        super().__init__()
+        self.norm      = nn.RMSNorm(hidden_size, eps=rms_norm_eps)
+        self.up_proj   = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        nn.init.zeros_(self.down_proj.weight)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x + self.down_proj(F.silu(self.up_proj(self.norm(x))))
+
+
+class ParallelMLPBlockWrapper(nn.Module):
+    """Wraps a transformer decoder layer, adding a parallel MLP adapter.
+
+    The adapter sees the same pre-block hidden states as the original block.
+    Its delta (adapter output minus input) is added to the original block's output,
+    so neither path depends on the other — true parallel execution.
+    """
+
+    def __init__(self, original_block: nn.Module, mlp_adapter: ParallelMLPAdapter):
+        super().__init__()
+        self.original_block = original_block
+        self.mlp_adapter = mlp_adapter  # name used by freeze filter: "mlp_adapter"
+
+    def __getattr__(self, name: str):
+        try:
+            return super().__getattr__(name)
+        except AttributeError:
+            return getattr(self.original_block, name)
+
+    def forward(self, hidden_states: torch.Tensor, *args, **kwargs):
+        block_outputs = self.original_block(hidden_states, *args, **kwargs)
+        out_hidden = block_outputs[0] if isinstance(block_outputs, tuple) else block_outputs
+
+        # adapter(x) returns x + delta; extract only the delta to add to block output
+        adapter_delta = self.mlp_adapter(hidden_states) - hidden_states
+        out_hidden = out_hidden + adapter_delta
+
+        if isinstance(block_outputs, tuple):
+            return (out_hidden,) + block_outputs[1:]
+        return out_hidden
+
+
+class UnifiedParallelMLPForCausalLM(PreTrainedModel):
+    config_class = ParallelMLPConfig
+
+    def __init__(self, config: ParallelMLPConfig, **kwargs):
+        super().__init__(config)
+        self.backbone = AutoModelForCausalLM.from_pretrained(
+            config.base_model_name_or_path,
+            torch_dtype=torch.bfloat16,
+            trust_remote_code=True,
+            attn_implementation="flash_attention_2",
+        )
+
+        backbone_dtype = next(self.backbone.parameters()).dtype
+        layers = self.backbone.model.layers
+        for pos in config.mlp_positions:
+            if 0 < pos <= len(layers):
+                original_layer = layers[pos - 1]
+                adapter = ParallelMLPAdapter(
+                    config.hidden_size,
+                    config.mlp_intermediate_size,
+                    config.rms_norm_eps,
+                ).to(dtype=backbone_dtype)
+                layers[pos - 1] = ParallelMLPBlockWrapper(original_layer, adapter)
+
+    def forward(self, *args, **kwargs):
+        return self.backbone(*args, **kwargs)
+
+    def generate(self, *args, **kwargs):
+        return self.backbone.generate(*args, **kwargs)

special_tokens_map.json

@@ -0,0 +1,25 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|vision_pad|>"
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c5ae00e602b8860cbd784ba82a8aa14e8feecec692e7076590d014d7b7fdafa
+size 11421896

tokenizer_config.json

@@ -0,0 +1,207 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|vision_pad|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}