Upload folder using huggingface_hub

.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

README.md

@@ -0,0 +1,151 @@
+---
+license: apache-2.0
+license_link: https://ai.google.dev/gemma/docs/gemma_4_license
+pipeline_tag: image-text-to-text
+library_name: transformers
+base_model: trohrbaugh/gemma-4-31b-it-heretic-ara
+base_model_relation: quantized
+tags:
+  - gemma
+  - gemma-4
+  - multimodal
+  - vision-language
+  - abliterated
+  - heretic
+  - ara
+  - uncensored
+  - decensored
+  - awq
+  - w4a16
+  - quantized
+  - compressed-tensors
+  - vllm
+---
+
+# Gemma-4-31B-it-abliterated-heretic-AWQ-W4A16
+
+AWQ W4A16 (group_size 128, symmetric) quantization of
+[trohrbaugh/gemma-4-31b-it-heretic-ara](https://huggingface.co/trohrbaugh/gemma-4-31b-it-heretic-ara)
+— a Heretic-ARA abliterated derivative of [google/gemma-4-31b-it](https://huggingface.co/google/gemma-4-31b-it).
+
+> ⚠️ **Decensored model.** Safety guardrails have been deliberately removed.
+> Research and experimentation only. See full disclaimer below.
+
+## Quantization Details
+| Parameter | Value |
+| :-------- | :---: |
+| **Method** | AWQ (Activation-aware Weight Quantization) |
+| **Scheme** | W4A16 (symmetric) |
+| **Weight Bits** | 4 |
+| **Activation Bits** | 16 |
+| **Group Size** | 128 |
+| **Format** | compressed-tensors |
+| **Calibration Dataset** | HuggingFaceH4/ultrachat_200k |
+| **Calibration Samples** | 256 |
+| **Max Sequence Length** | 2048 |
+| **Vision Tower** | Unquantized (full precision) |
+| **LM Head** | Unquantized (full precision) |
+| **Compatible Inference Engine** | vLLM (`vllm/vllm-openai:gemma4`) |
+
+## Quantization Notes
+
+- **All multimodal paths kept full precision**: Vision tower, audio tower, video
+  tower, multi-modal projector, and all modality-specific embedding and
+  projection layers are excluded from quantization. Only language-model linear
+  layers (attention Q/K/V/O and MLP gate/up/down) are quantized.
+- **LM head unquantized**: Standard practice to preserve output token
+  distribution quality at negligible size cost.
+- **v_proj → o_proj smoothing skipped**: llm-compressor reports incompatible
+  balance layer dimensions for `v_proj → o_proj` on this checkpoint across many
+  of Gemma 4's decoder blocks. Per-channel smoothing is skipped for that path,
+  matching the standard AWQ-for-GQA pattern.
+- **Hybrid attention and context window unchanged**: Quantization only touches
+  Linear layer weights; Gemma 4's interleaved local/global attention pattern
+  and 256K context capacity are structurally preserved. Actual long-context
+  quality at 4-bit has not been benchmarked.
+- **Multimodal input intact**: Text + image input works as in the base model.
+  Use the standard Gemma 4 chat template with image tokens placed before text.
+- **Full quantization recipe** is preserved in `recipe.yaml` in this repo for
+  reproducibility. It records the exact AWQ mappings, ignore patterns, and
+  scheme parameters applied.
+
+## Deployment
+
+Recommended inference with vLLM (Gemma 4 requires the `vllm/vllm-openai:gemma4` image):
+
+```bash
+vllm serve alonsoko/gemma-4-31b-it-abliterated-heretic-AWQ-W4A16 \
+    --trust-remote-code \
+    --tensor-parallel-size 1 \
+    --max-model-len 32768
+```
+
+Recommended sampling (per upstream Gemma 4 guidance):
+`temperature=1.0, top_p=0.95, top_k=64`
+
+To enable thinking mode, include the `<|think|>` token at the start of the system prompt.
+
+## Hardware Requirements
+
+Approximate VRAM for inference at this quantization (W4A16-G128):
+- **Weights** (quantized language model + unquantized vision tower, all in one safetensors): **~19 GB**
+- **KV cache** (per request, grows with context length): ~1-4 GB at 32K context, more at longer contexts
+- **Recommended**: Single 24 GB GPU (RTX 3090/4090, A10G) for standard context up to ~32K,
+  or single 48 GB GPU (L40S/A6000) for long context / batch serving
+- **Long context (128K+)**: 48-80 GB recommended due to KV cache growth
+
+---
+
+## ⚠️ Disclaimer
+
+**This model is intended for research, experimentation, and testing purposes only.**
+
+- This model may produce harmful, offensive, inappropriate, or otherwise objectionable content.
+- The abliteration process removes safety guardrails that were intentionally built into the original model.
+- Do **not** use this model in production systems, consumer-facing applications, or any context
+  where harmful outputs could cause real-world harm.
+- The authors and contributors of this toolkit bear no responsibility for any misuse of this model
+  or any harm caused by outputs generated by this model.
+- By using this model, you agree that you are solely responsible for ensuring its use complies
+  with all applicable laws and ethical guidelines.
+
+This model is shared purely for academic and technical exploration of model internals.
+
+## Abliteration
+
+Performed with [Heretic](https://github.com/p-e-w/heretic) v1.2.0+custom using the
+[Arbitrary-Rank Ablation (ARA)](https://github.com/p-e-w/heretic/pull/211) method.
+
+| Parameter | Value |
+| :-------- | :---: |
+| **start_layer_index** | 2 |
+| **end_layer_index** | 60 |
+| **preserve_good_behavior_weight** | 0.9920 |
+| **steer_bad_behavior_weight** | 0.0001 |
+| **overcorrect_relative_weight** | 0.4709 |
+| **neighbor_count** | 10 |
+
+### Performance
+
+| Metric | This model | Original [google/gemma-4-31b-it](https://huggingface.co/google/gemma-4-31b-it) |
+| :----- | :--------: | :---------------------------: |
+| **KL divergence** | 0.0120 | 0 *(by definition)* |
+| **Refusals** | 5/100 | 98/100 |
+
+*Measured on the unquantized heretic-ara base; AWQ is expected to preserve these closely but has not been separately benchmarked.*
+
+---
+
+## About the Base Model
+
+Original model: [google/gemma-4-31b-it](https://huggingface.co/google/gemma-4-31b-it)
+
+Gemma 4 31B is a dense multimodal model (text + image input, text output) with
+a 256K context window, native thinking-mode support, function calling, and
+strong performance on reasoning, coding, and vision benchmarks. See the
+[base model card](https://huggingface.co/google/gemma-4-31b-it) for architectural
+details, benchmark results, training data, and Google's responsible-use guidance.
+
+<div align="center">
+  <img src="https://ai.google.dev/gemma/images/gemma4_banner.png" alt="Gemma 4">
+</div>

chat_template.jinja

@@ -0,0 +1,266 @@
+{%- macro format_parameters(properties, required) -%}
+    {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
+    {%- set ns = namespace(found_first=false) -%}
+    {%- for key, value in properties | dictsort -%}
+        {%- set add_comma = false -%}
+        {%- if key not in standard_keys -%}
+            {%- if ns.found_first %},{% endif -%}
+            {%- set ns.found_first = true -%}
+            {{ key }}:{
+            {%- if value['description'] -%}
+                description:<|"|>{{ value['description'] }}<|"|>
+                {%- set add_comma = true -%}
+            {%- endif -%}
+            {%- if value['nullable'] %}
+                {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                nullable:true
+            {%- endif -%}
+            {%- if value['type'] | upper == 'STRING' -%}
+                {%- if value['enum'] -%}
+                    {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+                    enum:{{ format_argument(value['enum']) }}
+                {%- endif -%}
+            {%- elif value['type'] | upper == 'OBJECT' -%}
+                ,properties:{
+                {%- if value['properties'] is defined and value['properties'] is mapping -%}
+                    {{- format_parameters(value['properties'], value['required'] | default([])) -}}
+                {%- elif value is mapping -%}
+                    {{- format_parameters(value, value['required'] | default([])) -}}
+                {%- endif -%}
+                }
+                {%- if value['required'] -%}
+                    ,required:[
+                    {%- for item in value['required'] | default([]) -%}
+                        <|"|>{{- item -}}<|"|>
+                        {%- if not loop.last %},{% endif -%}
+                    {%- endfor -%}
+                    ]
+                {%- endif -%}
+            {%- elif value['type'] | upper == 'ARRAY' -%}
+                {%- if value['items'] is mapping and value['items'] -%}
+                    ,items:{
+                    {%- set ns_items = namespace(found_first=false) -%}
+                    {%- for item_key, item_value in value['items'] | dictsort -%}
+                        {%- if item_value is not none -%}
+                            {%- if ns_items.found_first %},{% endif -%}
+                            {%- set ns_items.found_first = true -%}
+                            {%- if item_key == 'properties' -%}
+                                properties:{
+                                {%- if item_value is mapping -%}
+                                    {{- format_parameters(item_value, value['items']['required'] | default([])) -}}
+                                {%- endif -%}
+                                }
+                            {%- elif item_key == 'required' -%}
+                                required:[
+                                {%- for req_item in item_value -%}
+                                    <|"|>{{- req_item -}}<|"|>
+                                    {%- if not loop.last %},{% endif -%}
+                                {%- endfor -%}
+                                ]
+                            {%- elif item_key == 'type' -%}
+                                {%- if item_value is string -%}
+                                    type:{{ format_argument(item_value | upper) }}
+                                {%- else -%}
+                                    type:{{ format_argument(item_value | map('upper') | list) }}
+                                {%- endif -%}
+                            {%- else -%}
+                                {{ item_key }}:{{ format_argument(item_value) }}
+                            {%- endif -%}
+                        {%- endif -%}
+                    {%- endfor -%}
+                    }
+                {%- endif -%}
+            {%- endif -%}
+            {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}
+            type:<|"|>{{ value['type'] | upper }}<|"|>}
+        {%- endif -%}
+    {%- endfor -%}
+{%- endmacro -%}
+{%- macro format_function_declaration(tool_data) -%}
+    declaration:{{- tool_data['function']['name'] -}}{description:<|"|>{{- tool_data['function']['description'] -}}<|"|>
+    {%- set params = tool_data['function']['parameters'] -%}
+    {%- if params -%}
+        ,parameters:{
+        {%- if params['properties'] -%}
+            properties:{ {{- format_parameters(params['properties'], params['required']) -}} },
+        {%- endif -%}
+        {%- if params['required'] -%}
+            required:[
+            {%- for item in params['required'] -%}
+                <|"|>{{- item -}}<|"|>
+                {{- ',' if not loop.last -}}
+            {%- endfor -%}
+            ],
+        {%- endif -%}
+        {%- if params['type'] -%}
+            type:<|"|>{{- params['type'] | upper -}}<|"|>}
+        {%- endif -%}
+    {%- endif -%}
+    {%- if 'response' in tool_data['function'] -%}
+        {%- set response_declaration = tool_data['function']['response'] -%}
+        ,response:{
+        {%- if response_declaration['description'] -%}
+            description:<|"|>{{- response_declaration['description'] -}}<|"|>,
+        {%- endif -%}
+        {%- if response_declaration['type'] | upper == 'OBJECT' -%}
+            type:<|"|>{{- response_declaration['type'] | upper -}}<|"|>}
+        {%- endif -%}
+    {%- endif -%}
+    }
+{%- endmacro -%}
+{%- macro format_argument(argument, escape_keys=True) -%}
+    {%- if argument is string -%}
+        {{- '<|"|>' + argument + '<|"|>' -}}
+    {%- elif argument is boolean -%}
+        {{- 'true' if argument else 'false' -}}
+    {%- elif argument is mapping -%}
+        {{- '{' -}}
+        {%- set ns = namespace(found_first=false) -%}
+        {%- for key, value in argument | dictsort -%}
+            {%- if ns.found_first %},{% endif -%}
+            {%- set ns.found_first = true -%}
+            {%- if escape_keys -%}
+                {{- '<|"|>' + key + '<|"|>' -}}
+            {%- else -%}
+                {{- key -}}
+            {%- endif -%}
+            :{{- format_argument(value, escape_keys=escape_keys) -}}
+        {%- endfor -%}
+        {{- '}' -}}
+    {%- elif argument is sequence -%}
+        {{- '[' -}}
+        {%- for item in argument -%}
+            {{- format_argument(item, escape_keys=escape_keys) -}}
+            {%- if not loop.last %},{% endif -%}
+        {%- endfor -%}
+        {{- ']' -}}
+    {%- else -%}
+        {{- argument -}}
+    {%- endif -%}
+{%- endmacro -%}
+{%- macro strip_thinking(text) -%}
+    {%- set ns = namespace(result='') -%}
+    {%- for part in text.split('<channel|>') -%}
+        {%- if '<|channel>' in part -%}
+            {%- set ns.result = ns.result + part.split('<|channel>')[0] -%}
+        {%- else -%}
+            {%- set ns.result = ns.result + part -%}
+        {%- endif -%}
+    {%- endfor -%}
+    {{- ns.result | trim -}}
+{%- endmacro -%}
+
+{%- set ns = namespace(prev_message_type=None) -%}
+{%- set loop_messages = messages -%}
+{{ bos_token }}
+{#- Handle System/Tool Definitions Block -#}
+{%- if (enable_thinking is defined and enable_thinking) or tools or messages[0]['role'] in ['system', 'developer'] -%}
+    {{- '<|turn>system\n' -}}
+
+    {#- Inject Thinking token at the very top of the FIRST system turn -#}
+    {%- if enable_thinking is defined and enable_thinking -%}
+        {{- '<|think|>' -}}
+        {%- set ns.prev_message_type = 'think' -%}
+    {%- endif -%}
+
+    {%- if messages[0]['role'] in ['system', 'developer'] -%}
+        {{- messages[0]['content'] | trim -}}
+        {%- set loop_messages = messages[1:] -%}
+    {%- endif -%}
+
+    {%- if tools -%}
+        {%- for tool in tools %}
+            {{- '<|tool>' -}}
+            {{- format_function_declaration(tool) | trim -}}
+            {{- '<tool|>' -}}
+        {%- endfor %}
+        {%- set ns.prev_message_type = 'tool' -%}
+    {%- endif -%}
+
+    {{- '<turn|>\n' -}}
+{%- endif %}
+
+{#- Loop through messages -#}
+{%- for message in loop_messages -%}
+    {%- set ns.prev_message_type = None -%}
+    {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}
+        {{- '<|turn>' + role + '\n' }}
+
+            {%- if message['tool_calls'] -%}
+                {%- for tool_call in message['tool_calls'] -%}
+                    {%- set function = tool_call['function'] -%}
+                    {{- '<|tool_call>call:' + function['name'] + '{' -}}
+                    {%- if function['arguments'] is mapping -%}
+                        {%- set ns_args = namespace(found_first=false) -%}
+                        {%- for key, value in function['arguments'] | dictsort -%}
+                            {%- if ns_args.found_first %},{% endif -%}
+                            {%- set ns_args.found_first = true -%}
+                            {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
+                        {%- endfor -%}
+                    {%- elif function['arguments'] is string -%}
+                        {{- function['arguments'] -}}
+                    {%- endif -%}
+                    {{- '}<tool_call|>' -}}
+                {%- endfor -%}
+                {%- set ns.prev_message_type = 'tool_call' -%}
+            {%- endif -%}
+
+            {%- if message['tool_responses'] -%}
+                {#- Tool Response handling -#}
+                {%- for tool_response in message['tool_responses'] -%}
+                    {{- '<|tool_response>' -}}
+                    {%- if tool_response['response'] is mapping -%}
+                        {{- 'response:' + tool_response['name'] | default('unknown') + '{' -}}
+                        {%- for key, value in tool_response['response'] | dictsort -%}
+                            {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
+                            {%- if not loop.last %},{% endif -%}
+                        {%- endfor -%}
+                        {{- '}' -}}
+                    {%- else -%}
+                        {{- 'response:' + tool_response['name'] | default('unknown') + '{value:' + format_argument(tool_response['response'], escape_keys=False) + '}' -}}
+                    {%- endif -%}
+                    {{- '<tool_response|>' -}}
+                {%- endfor -%}
+                {%- set ns.prev_message_type = 'tool_response' -%}
+            {%- endif -%}
+
+            {%- if message['content'] is string -%}
+                {%- if role == 'model' -%}
+                    {{- strip_thinking(message['content']) -}}
+                {%- else -%}
+                    {{- message['content'] | trim -}}
+                {%- endif -%}
+            {%- elif message['content'] is sequence -%}
+                {%- for item in message['content'] -%}
+                    {%- if item['type'] == 'text' -%}
+                        {%- if role == 'model' -%}
+                            {{- strip_thinking(item['text']) -}}
+                        {%- else -%}
+                            {{- item['text'] | trim -}}
+                        {%- endif -%}
+                    {%- elif item['type'] == 'image' -%}
+                        {{- '\n\n<|image|>\n\n' -}}
+                        {%- set ns.prev_message_type = 'image' -%}
+                    {%- elif item['type'] == 'audio' -%}
+                        {{- '<|audio|>' -}}
+                        {%- set ns.prev_message_type = 'audio' -%}
+                    {%- elif item['type'] == 'video' -%}
+                        {{- '\n\n<|video|>\n\n' -}}
+                        {%- set ns.prev_message_type = 'video' -%}
+                    {%- endif -%}
+                {%- endfor -%}
+            {%- endif -%}
+
+        {%- if not (message['tool_responses'] and not message['content']) -%}
+            {{- '<turn|>\n' -}}
+        {%- endif -%}
+{%- endfor -%}
+
+{%- if add_generation_prompt -%}
+    {%- if ns.prev_message_type != 'tool_response' -%}
+        {{- '<|turn>model\n' -}}
+    {%- endif -%}
+    {%- if not enable_thinking | default(false) -%}
+        {{- '<|channel>thought\n<channel|>' -}}
+    {%- endif -%}
+{%- endif -%}

config.json

@@ -0,0 +1,349 @@
+{
+  "architectures": [
+    "Gemma4ForCausalLM"
+  ],
+  "boa_token_id": 256000,
+  "boi_token_id": 255999,
+  "dtype": "bfloat16",
+  "eoa_token_id": 258883,
+  "eoa_token_index": 258883,
+  "eoi_token_id": 258882,
+  "eos_token_id": 1,
+  "image_token_id": 258880,
+  "initializer_range": 0.02,
+  "model_type": "gemma4_text",
+  "quantization_config": {
+    "config_groups": {
+      "group_0": {
+        "format": "pack-quantized",
+        "input_activations": null,
+        "output_activations": null,
+        "targets": [
+          "Linear"
+        ],
+        "weights": {
+          "actorder": null,
+          "block_structure": null,
+          "dynamic": false,
+          "group_size": 128,
+          "num_bits": 4,
+          "observer": "memoryless_minmax",
+          "observer_kwargs": {},
+          "scale_dtype": null,
+          "strategy": "group",
+          "symmetric": true,
+          "type": "int",
+          "zp_dtype": null
+        }
+      }
+    },
+    "format": "pack-quantized",
+    "global_compression_ratio": null,
+    "ignore": [
+      "model.vision_tower.patch_embedder.input_proj",
+      "model.vision_tower.encoder.layers.0.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.0.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.0.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.0.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.0.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.0.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.0.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.1.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.1.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.1.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.1.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.1.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.1.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.1.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.2.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.2.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.2.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.2.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.2.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.2.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.2.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.3.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.3.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.3.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.3.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.3.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.3.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.3.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.4.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.4.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.4.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.4.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.4.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.4.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.4.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.5.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.5.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.5.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.5.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.5.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.5.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.5.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.6.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.6.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.6.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.6.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.6.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.6.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.6.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.7.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.7.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.7.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.7.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.7.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.7.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.7.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.8.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.8.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.8.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.8.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.8.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.8.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.8.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.9.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.9.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.9.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.9.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.9.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.9.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.9.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.10.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.10.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.10.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.10.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.10.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.10.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.10.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.11.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.11.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.11.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.11.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.11.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.11.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.11.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.12.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.12.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.12.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.12.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.12.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.12.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.12.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.13.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.13.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.13.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.13.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.13.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.13.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.13.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.14.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.14.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.14.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.14.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.14.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.14.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.14.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.15.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.15.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.15.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.15.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.15.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.15.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.15.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.16.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.16.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.16.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.16.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.16.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.16.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.16.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.17.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.17.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.17.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.17.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.17.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.17.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.17.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.18.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.18.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.18.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.18.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.18.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.18.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.18.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.19.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.19.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.19.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.19.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.19.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.19.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.19.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.20.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.20.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.20.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.20.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.20.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.20.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.20.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.21.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.21.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.21.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.21.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.21.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.21.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.21.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.22.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.22.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.22.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.22.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.22.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.22.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.22.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.23.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.23.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.23.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.23.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.23.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.23.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.23.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.24.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.24.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.24.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.24.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.24.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.24.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.24.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.25.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.25.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.25.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.25.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.25.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.25.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.25.mlp.down_proj.linear",
+      "model.vision_tower.encoder.layers.26.self_attn.q_proj.linear",
+      "model.vision_tower.encoder.layers.26.self_attn.k_proj.linear",
+      "model.vision_tower.encoder.layers.26.self_attn.v_proj.linear",
+      "model.vision_tower.encoder.layers.26.self_attn.o_proj.linear",
+      "model.vision_tower.encoder.layers.26.mlp.gate_proj.linear",
+      "model.vision_tower.encoder.layers.26.mlp.up_proj.linear",
+      "model.vision_tower.encoder.layers.26.mlp.down_proj.linear",
+      "model.embed_vision.embedding_projection",
+      "lm_head"
+    ],
+    "kv_cache_scheme": null,
+    "quant_method": "compressed-tensors",
+    "quantization_status": "compressed",
+    "sparsity_config": {},
+    "transform_config": {},
+    "version": "0.15.1.dev3+g7946613"
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "5.5.1",
+  "video_token_id": 258884,
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attention_k_eq_v": true,
+  "bos_token_id": 2,
+  "enable_moe_block": false,
+  "expert_intermediate_size": null,
+  "final_logit_softcapping": 30.0,
+  "global_head_dim": 512,
+  "head_dim": 256,
+  "hidden_activation": "gelu_pytorch_tanh",
+  "hidden_size": 5376,
+  "hidden_size_per_layer_input": 0,
+  "intermediate_size": 21504,
+  "layer_types": [
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 262144,
+  "moe_intermediate_size": null,
+  "num_attention_heads": 32,
+  "num_experts": null,
+  "num_global_key_value_heads": 4,
+  "num_hidden_layers": 60,
+  "num_key_value_heads": 16,
+  "num_kv_shared_layers": 0,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "full_attention": {
+      "partial_rotary_factor": 0.25,
+      "rope_theta": 1000000.0,
+      "rope_type": "proportional"
+    },
+    "sliding_attention": {
+      "rope_theta": 10000.0,
+      "rope_type": "default"
+    }
+  },
+  "sliding_window": 1024,
+  "top_k_experts": null,
+  "use_bidirectional_attention": "vision",
+  "use_cache": true,
+  "use_double_wide_mlp": false,
+  "vocab_size": 262144,
+  "vocab_size_per_layer_input": 262144
+}

generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 2,
+  "do_sample": true,
+  "eos_token_id": [
+    1,
+    106,
+    50
+  ],
+  "pad_token_id": 0,
+  "temperature": 1.0,
+  "top_k": 64,
+  "top_p": 0.95,
+  "transformers_version": "5.5.1"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:413e6592218401afba67a946856f35d318db16804656a0554b4ef9f29d4756a0
+size 17934788136

recipe.yaml

@@ -0,0 +1,573 @@
+default_stage:
+  default_modifiers:
+    AWQModifier:
+      targets: [Linear]
+      ignore: [lm_head, 're:.*vision_tower.*', 're:.*audio_tower.*', 're:.*video_tower.*',
+        're:.*multi_modal_projector.*', 're:.*embed_vision.*', 're:.*embed_audio.*', 're:.*embed_video.*',
+        're:.*image.*projection.*', 're:.*video.*projection.*', 're:.*audio.*projection.*']
+      scheme: W4A16
+      bypass_divisibility_checks: false
+      mappings:
+      - smooth_layer: model.language_model.layers.0.input_layernorm
+        balance_layers: [model.language_model.layers.0.self_attn.q_proj, model.language_model.layers.0.self_attn.k_proj,
+          model.language_model.layers.0.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.0.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.0.mlp.gate_proj, model.language_model.layers.0.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.0.mlp.up_proj
+        balance_layers: [model.language_model.layers.0.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.1.input_layernorm
+        balance_layers: [model.language_model.layers.1.self_attn.q_proj, model.language_model.layers.1.self_attn.k_proj,
+          model.language_model.layers.1.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.1.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.1.mlp.gate_proj, model.language_model.layers.1.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.1.mlp.up_proj
+        balance_layers: [model.language_model.layers.1.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.2.input_layernorm
+        balance_layers: [model.language_model.layers.2.self_attn.q_proj, model.language_model.layers.2.self_attn.k_proj,
+          model.language_model.layers.2.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.2.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.2.mlp.gate_proj, model.language_model.layers.2.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.2.mlp.up_proj
+        balance_layers: [model.language_model.layers.2.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.3.input_layernorm
+        balance_layers: [model.language_model.layers.3.self_attn.q_proj, model.language_model.layers.3.self_attn.k_proj,
+          model.language_model.layers.3.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.3.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.3.mlp.gate_proj, model.language_model.layers.3.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.3.mlp.up_proj
+        balance_layers: [model.language_model.layers.3.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.4.input_layernorm
+        balance_layers: [model.language_model.layers.4.self_attn.q_proj, model.language_model.layers.4.self_attn.k_proj,
+          model.language_model.layers.4.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.4.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.4.mlp.gate_proj, model.language_model.layers.4.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.4.mlp.up_proj
+        balance_layers: [model.language_model.layers.4.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.5.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.5.mlp.gate_proj, model.language_model.layers.5.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.5.mlp.up_proj
+        balance_layers: [model.language_model.layers.5.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.6.input_layernorm
+        balance_layers: [model.language_model.layers.6.self_attn.q_proj, model.language_model.layers.6.self_attn.k_proj,
+          model.language_model.layers.6.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.6.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.6.mlp.gate_proj, model.language_model.layers.6.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.6.mlp.up_proj
+        balance_layers: [model.language_model.layers.6.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.7.input_layernorm
+        balance_layers: [model.language_model.layers.7.self_attn.q_proj, model.language_model.layers.7.self_attn.k_proj,
+          model.language_model.layers.7.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.7.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.7.mlp.gate_proj, model.language_model.layers.7.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.7.mlp.up_proj
+        balance_layers: [model.language_model.layers.7.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.8.input_layernorm
+        balance_layers: [model.language_model.layers.8.self_attn.q_proj, model.language_model.layers.8.self_attn.k_proj,
+          model.language_model.layers.8.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.8.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.8.mlp.gate_proj, model.language_model.layers.8.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.8.mlp.up_proj
+        balance_layers: [model.language_model.layers.8.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.9.input_layernorm
+        balance_layers: [model.language_model.layers.9.self_attn.q_proj, model.language_model.layers.9.self_attn.k_proj,
+          model.language_model.layers.9.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.9.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.9.mlp.gate_proj, model.language_model.layers.9.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.9.mlp.up_proj
+        balance_layers: [model.language_model.layers.9.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.10.input_layernorm
+        balance_layers: [model.language_model.layers.10.self_attn.q_proj, model.language_model.layers.10.self_attn.k_proj,
+          model.language_model.layers.10.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.10.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.10.mlp.gate_proj, model.language_model.layers.10.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.10.mlp.up_proj
+        balance_layers: [model.language_model.layers.10.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.11.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.11.mlp.gate_proj, model.language_model.layers.11.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.11.mlp.up_proj
+        balance_layers: [model.language_model.layers.11.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.12.input_layernorm
+        balance_layers: [model.language_model.layers.12.self_attn.q_proj, model.language_model.layers.12.self_attn.k_proj,
+          model.language_model.layers.12.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.12.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.12.mlp.gate_proj, model.language_model.layers.12.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.12.mlp.up_proj
+        balance_layers: [model.language_model.layers.12.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.13.input_layernorm
+        balance_layers: [model.language_model.layers.13.self_attn.q_proj, model.language_model.layers.13.self_attn.k_proj,
+          model.language_model.layers.13.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.13.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.13.mlp.gate_proj, model.language_model.layers.13.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.13.mlp.up_proj
+        balance_layers: [model.language_model.layers.13.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.14.input_layernorm
+        balance_layers: [model.language_model.layers.14.self_attn.q_proj, model.language_model.layers.14.self_attn.k_proj,
+          model.language_model.layers.14.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.14.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.14.mlp.gate_proj, model.language_model.layers.14.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.14.mlp.up_proj
+        balance_layers: [model.language_model.layers.14.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.15.input_layernorm
+        balance_layers: [model.language_model.layers.15.self_attn.q_proj, model.language_model.layers.15.self_attn.k_proj,
+          model.language_model.layers.15.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.15.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.15.mlp.gate_proj, model.language_model.layers.15.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.15.mlp.up_proj
+        balance_layers: [model.language_model.layers.15.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.16.input_layernorm
+        balance_layers: [model.language_model.layers.16.self_attn.q_proj, model.language_model.layers.16.self_attn.k_proj,
+          model.language_model.layers.16.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.16.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.16.mlp.gate_proj, model.language_model.layers.16.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.16.mlp.up_proj
+        balance_layers: [model.language_model.layers.16.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.17.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.17.mlp.gate_proj, model.language_model.layers.17.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.17.mlp.up_proj
+        balance_layers: [model.language_model.layers.17.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.18.input_layernorm
+        balance_layers: [model.language_model.layers.18.self_attn.q_proj, model.language_model.layers.18.self_attn.k_proj,
+          model.language_model.layers.18.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.18.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.18.mlp.gate_proj, model.language_model.layers.18.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.18.mlp.up_proj
+        balance_layers: [model.language_model.layers.18.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.19.input_layernorm
+        balance_layers: [model.language_model.layers.19.self_attn.q_proj, model.language_model.layers.19.self_attn.k_proj,
+          model.language_model.layers.19.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.19.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.19.mlp.gate_proj, model.language_model.layers.19.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.19.mlp.up_proj
+        balance_layers: [model.language_model.layers.19.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.20.input_layernorm
+        balance_layers: [model.language_model.layers.20.self_attn.q_proj, model.language_model.layers.20.self_attn.k_proj,
+          model.language_model.layers.20.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.20.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.20.mlp.gate_proj, model.language_model.layers.20.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.20.mlp.up_proj
+        balance_layers: [model.language_model.layers.20.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.21.input_layernorm
+        balance_layers: [model.language_model.layers.21.self_attn.q_proj, model.language_model.layers.21.self_attn.k_proj,
+          model.language_model.layers.21.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.21.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.21.mlp.gate_proj, model.language_model.layers.21.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.21.mlp.up_proj
+        balance_layers: [model.language_model.layers.21.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.22.input_layernorm
+        balance_layers: [model.language_model.layers.22.self_attn.q_proj, model.language_model.layers.22.self_attn.k_proj,
+          model.language_model.layers.22.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.22.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.22.mlp.gate_proj, model.language_model.layers.22.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.22.mlp.up_proj
+        balance_layers: [model.language_model.layers.22.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.23.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.23.mlp.gate_proj, model.language_model.layers.23.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.23.mlp.up_proj
+        balance_layers: [model.language_model.layers.23.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.24.input_layernorm
+        balance_layers: [model.language_model.layers.24.self_attn.q_proj, model.language_model.layers.24.self_attn.k_proj,
+          model.language_model.layers.24.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.24.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.24.mlp.gate_proj, model.language_model.layers.24.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.24.mlp.up_proj
+        balance_layers: [model.language_model.layers.24.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.25.input_layernorm
+        balance_layers: [model.language_model.layers.25.self_attn.q_proj, model.language_model.layers.25.self_attn.k_proj,
+          model.language_model.layers.25.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.25.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.25.mlp.gate_proj, model.language_model.layers.25.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.25.mlp.up_proj
+        balance_layers: [model.language_model.layers.25.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.26.input_layernorm
+        balance_layers: [model.language_model.layers.26.self_attn.q_proj, model.language_model.layers.26.self_attn.k_proj,
+          model.language_model.layers.26.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.26.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.26.mlp.gate_proj, model.language_model.layers.26.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.26.mlp.up_proj
+        balance_layers: [model.language_model.layers.26.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.27.input_layernorm
+        balance_layers: [model.language_model.layers.27.self_attn.q_proj, model.language_model.layers.27.self_attn.k_proj,
+          model.language_model.layers.27.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.27.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.27.mlp.gate_proj, model.language_model.layers.27.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.27.mlp.up_proj
+        balance_layers: [model.language_model.layers.27.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.28.input_layernorm
+        balance_layers: [model.language_model.layers.28.self_attn.q_proj, model.language_model.layers.28.self_attn.k_proj,
+          model.language_model.layers.28.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.28.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.28.mlp.gate_proj, model.language_model.layers.28.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.28.mlp.up_proj
+        balance_layers: [model.language_model.layers.28.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.29.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.29.mlp.gate_proj, model.language_model.layers.29.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.29.mlp.up_proj
+        balance_layers: [model.language_model.layers.29.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.30.input_layernorm
+        balance_layers: [model.language_model.layers.30.self_attn.q_proj, model.language_model.layers.30.self_attn.k_proj,
+          model.language_model.layers.30.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.30.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.30.mlp.gate_proj, model.language_model.layers.30.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.30.mlp.up_proj
+        balance_layers: [model.language_model.layers.30.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.31.input_layernorm
+        balance_layers: [model.language_model.layers.31.self_attn.q_proj, model.language_model.layers.31.self_attn.k_proj,
+          model.language_model.layers.31.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.31.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.31.mlp.gate_proj, model.language_model.layers.31.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.31.mlp.up_proj
+        balance_layers: [model.language_model.layers.31.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.32.input_layernorm
+        balance_layers: [model.language_model.layers.32.self_attn.q_proj, model.language_model.layers.32.self_attn.k_proj,
+          model.language_model.layers.32.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.32.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.32.mlp.gate_proj, model.language_model.layers.32.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.32.mlp.up_proj
+        balance_layers: [model.language_model.layers.32.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.33.input_layernorm
+        balance_layers: [model.language_model.layers.33.self_attn.q_proj, model.language_model.layers.33.self_attn.k_proj,
+          model.language_model.layers.33.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.33.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.33.mlp.gate_proj, model.language_model.layers.33.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.33.mlp.up_proj
+        balance_layers: [model.language_model.layers.33.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.34.input_layernorm
+        balance_layers: [model.language_model.layers.34.self_attn.q_proj, model.language_model.layers.34.self_attn.k_proj,
+          model.language_model.layers.34.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.34.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.34.mlp.gate_proj, model.language_model.layers.34.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.34.mlp.up_proj
+        balance_layers: [model.language_model.layers.34.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.35.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.35.mlp.gate_proj, model.language_model.layers.35.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.35.mlp.up_proj
+        balance_layers: [model.language_model.layers.35.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.36.input_layernorm
+        balance_layers: [model.language_model.layers.36.self_attn.q_proj, model.language_model.layers.36.self_attn.k_proj,
+          model.language_model.layers.36.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.36.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.36.mlp.gate_proj, model.language_model.layers.36.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.36.mlp.up_proj
+        balance_layers: [model.language_model.layers.36.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.37.input_layernorm
+        balance_layers: [model.language_model.layers.37.self_attn.q_proj, model.language_model.layers.37.self_attn.k_proj,
+          model.language_model.layers.37.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.37.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.37.mlp.gate_proj, model.language_model.layers.37.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.37.mlp.up_proj
+        balance_layers: [model.language_model.layers.37.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.38.input_layernorm
+        balance_layers: [model.language_model.layers.38.self_attn.q_proj, model.language_model.layers.38.self_attn.k_proj,
+          model.language_model.layers.38.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.38.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.38.mlp.gate_proj, model.language_model.layers.38.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.38.mlp.up_proj
+        balance_layers: [model.language_model.layers.38.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.39.input_layernorm
+        balance_layers: [model.language_model.layers.39.self_attn.q_proj, model.language_model.layers.39.self_attn.k_proj,
+          model.language_model.layers.39.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.39.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.39.mlp.gate_proj, model.language_model.layers.39.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.39.mlp.up_proj
+        balance_layers: [model.language_model.layers.39.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.40.input_layernorm
+        balance_layers: [model.language_model.layers.40.self_attn.q_proj, model.language_model.layers.40.self_attn.k_proj,
+          model.language_model.layers.40.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.40.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.40.mlp.gate_proj, model.language_model.layers.40.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.40.mlp.up_proj
+        balance_layers: [model.language_model.layers.40.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.41.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.41.mlp.gate_proj, model.language_model.layers.41.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.41.mlp.up_proj
+        balance_layers: [model.language_model.layers.41.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.42.input_layernorm
+        balance_layers: [model.language_model.layers.42.self_attn.q_proj, model.language_model.layers.42.self_attn.k_proj,
+          model.language_model.layers.42.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.42.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.42.mlp.gate_proj, model.language_model.layers.42.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.42.mlp.up_proj
+        balance_layers: [model.language_model.layers.42.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.43.input_layernorm
+        balance_layers: [model.language_model.layers.43.self_attn.q_proj, model.language_model.layers.43.self_attn.k_proj,
+          model.language_model.layers.43.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.43.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.43.mlp.gate_proj, model.language_model.layers.43.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.43.mlp.up_proj
+        balance_layers: [model.language_model.layers.43.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.44.input_layernorm
+        balance_layers: [model.language_model.layers.44.self_attn.q_proj, model.language_model.layers.44.self_attn.k_proj,
+          model.language_model.layers.44.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.44.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.44.mlp.gate_proj, model.language_model.layers.44.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.44.mlp.up_proj
+        balance_layers: [model.language_model.layers.44.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.45.input_layernorm
+        balance_layers: [model.language_model.layers.45.self_attn.q_proj, model.language_model.layers.45.self_attn.k_proj,
+          model.language_model.layers.45.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.45.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.45.mlp.gate_proj, model.language_model.layers.45.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.45.mlp.up_proj
+        balance_layers: [model.language_model.layers.45.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.46.input_layernorm
+        balance_layers: [model.language_model.layers.46.self_attn.q_proj, model.language_model.layers.46.self_attn.k_proj,
+          model.language_model.layers.46.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.46.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.46.mlp.gate_proj, model.language_model.layers.46.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.46.mlp.up_proj
+        balance_layers: [model.language_model.layers.46.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.47.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.47.mlp.gate_proj, model.language_model.layers.47.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.47.mlp.up_proj
+        balance_layers: [model.language_model.layers.47.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.48.input_layernorm
+        balance_layers: [model.language_model.layers.48.self_attn.q_proj, model.language_model.layers.48.self_attn.k_proj,
+          model.language_model.layers.48.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.48.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.48.mlp.gate_proj, model.language_model.layers.48.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.48.mlp.up_proj
+        balance_layers: [model.language_model.layers.48.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.49.input_layernorm
+        balance_layers: [model.language_model.layers.49.self_attn.q_proj, model.language_model.layers.49.self_attn.k_proj,
+          model.language_model.layers.49.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.49.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.49.mlp.gate_proj, model.language_model.layers.49.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.49.mlp.up_proj
+        balance_layers: [model.language_model.layers.49.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.50.input_layernorm
+        balance_layers: [model.language_model.layers.50.self_attn.q_proj, model.language_model.layers.50.self_attn.k_proj,
+          model.language_model.layers.50.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.50.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.50.mlp.gate_proj, model.language_model.layers.50.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.50.mlp.up_proj
+        balance_layers: [model.language_model.layers.50.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.51.input_layernorm
+        balance_layers: [model.language_model.layers.51.self_attn.q_proj, model.language_model.layers.51.self_attn.k_proj,
+          model.language_model.layers.51.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.51.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.51.mlp.gate_proj, model.language_model.layers.51.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.51.mlp.up_proj
+        balance_layers: [model.language_model.layers.51.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.52.input_layernorm
+        balance_layers: [model.language_model.layers.52.self_attn.q_proj, model.language_model.layers.52.self_attn.k_proj,
+          model.language_model.layers.52.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.52.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.52.mlp.gate_proj, model.language_model.layers.52.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.52.mlp.up_proj
+        balance_layers: [model.language_model.layers.52.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.53.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.53.mlp.gate_proj, model.language_model.layers.53.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.53.mlp.up_proj
+        balance_layers: [model.language_model.layers.53.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.54.input_layernorm
+        balance_layers: [model.language_model.layers.54.self_attn.q_proj, model.language_model.layers.54.self_attn.k_proj,
+          model.language_model.layers.54.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.54.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.54.mlp.gate_proj, model.language_model.layers.54.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.54.mlp.up_proj
+        balance_layers: [model.language_model.layers.54.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.55.input_layernorm
+        balance_layers: [model.language_model.layers.55.self_attn.q_proj, model.language_model.layers.55.self_attn.k_proj,
+          model.language_model.layers.55.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.55.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.55.mlp.gate_proj, model.language_model.layers.55.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.55.mlp.up_proj
+        balance_layers: [model.language_model.layers.55.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.56.input_layernorm
+        balance_layers: [model.language_model.layers.56.self_attn.q_proj, model.language_model.layers.56.self_attn.k_proj,
+          model.language_model.layers.56.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.56.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.56.mlp.gate_proj, model.language_model.layers.56.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.56.mlp.up_proj
+        balance_layers: [model.language_model.layers.56.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.57.input_layernorm
+        balance_layers: [model.language_model.layers.57.self_attn.q_proj, model.language_model.layers.57.self_attn.k_proj,
+          model.language_model.layers.57.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.57.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.57.mlp.gate_proj, model.language_model.layers.57.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.57.mlp.up_proj
+        balance_layers: [model.language_model.layers.57.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.58.input_layernorm
+        balance_layers: [model.language_model.layers.58.self_attn.q_proj, model.language_model.layers.58.self_attn.k_proj,
+          model.language_model.layers.58.self_attn.v_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.58.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.58.mlp.gate_proj, model.language_model.layers.58.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.58.mlp.up_proj
+        balance_layers: [model.language_model.layers.58.mlp.down_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.59.pre_feedforward_layernorm
+        balance_layers: [model.language_model.layers.59.mlp.gate_proj, model.language_model.layers.59.mlp.up_proj]
+        activation_hook_target: null
+      - smooth_layer: model.language_model.layers.59.mlp.up_proj
+        balance_layers: [model.language_model.layers.59.mlp.down_proj]
+        activation_hook_target: null
+      offload_device: cpu
+      duo_scaling: true
+      n_grid: 20

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cc9316e9b12bc8f317dd5b3fe2af39d4d314b71d59bf133c3f6d198736683ea
+size 32169892

tokenizer_config.json

@@ -0,0 +1,97 @@
+{
+  "audio_token": "<|audio|>",
+  "backend": "tokenizers",
+  "boa_token": "<|audio>",
+  "boi_token": "<|image>",
+  "bos_token": "<bos>",
+  "eoa_token": "<audio|>",
+  "eoc_token": "<channel|>",
+  "eoi_token": "<image|>",
+  "eos_token": "<eos>",
+  "eot_token": "<turn|>",
+  "escape_token": "<|\"|>",
+  "etc_token": "<tool_call|>",
+  "etd_token": "<tool|>",
+  "etr_token": "<tool_response|>",
+  "extra_special_tokens": [
+    "<|video|>"
+  ],
+  "image_token": "<|image|>",
+  "is_local": false,
+  "mask_token": "<mask>",
+  "max_length": null,
+  "model_max_length": 1000000000000000019884624838656,
+  "model_specific_special_tokens": {
+    "audio_token": "<|audio|>",
+    "boa_token": "<|audio>",
+    "boi_token": "<|image>",
+    "eoa_token": "<audio|>",
+    "eoc_token": "<channel|>",
+    "eoi_token": "<image|>",
+    "eot_token": "<turn|>",
+    "escape_token": "<|\"|>",
+    "etc_token": "<tool_call|>",
+    "etd_token": "<tool|>",
+    "etr_token": "<tool_response|>",
+    "image_token": "<|image|>",
+    "soc_token": "<|channel>",
+    "sot_token": "<|turn>",
+    "stc_token": "<|tool_call>",
+    "std_token": "<|tool>",
+    "str_token": "<|tool_response>",
+    "think_token": "<|think|>"
+  },
+  "pad_to_multiple_of": null,
+  "pad_token": "<pad>",
+  "pad_token_type_id": 0,
+  "padding_side": "left",
+  "response_schema": {
+    "properties": {
+      "content": {
+        "type": "string"
+      },
+      "role": {
+        "const": "assistant"
+      },
+      "thinking": {
+        "type": "string"
+      },
+      "tool_calls": {
+        "items": {
+          "properties": {
+            "function": {
+              "properties": {
+                "arguments": {
+                  "additionalProperties": {},
+                  "type": "object",
+                  "x-parser": "gemma4-tool-call"
+                },
+                "name": {
+                  "type": "string"
+                }
+              },
+              "type": "object",
+              "x-regex": "call\\:(?P<name>\\w+)(?P<arguments>\\{.*\\})"
+            },
+            "type": {
+              "const": "function"
+            }
+          },
+          "type": "object"
+        },
+        "type": "array",
+        "x-regex-iterator": "<\\|tool_call>(.*?)<tool_call\\|>"
+      }
+    },
+    "type": "object",
+    "x-regex": "(\\<\\|channel\\>thought\\n(?P<thinking>.*?)\\<channel\\|\\>)?(?P<content>(?:(?!\\<\\|tool_call\\>)(?!\\<turn\\|\\>).)+)?(?P<tool_calls>\\<\\|tool_call\\>.*\\<tool_call\\|\\>)?(?:\\<turn\\|\\>)?"
+  },
+  "soc_token": "<|channel>",
+  "sot_token": "<|turn>",
+  "stc_token": "<|tool_call>",
+  "std_token": "<|tool>",
+  "str_token": "<|tool_response>",
+  "think_token": "<|think|>",
+  "tokenizer_class": "GemmaTokenizer",
+  "unk_token": "<unk>"
+}