Model Upload

README.md

@@ -0,0 +1,174 @@
+---
+license: apache-2.0
+base_model:
+- stepfun-ai/step-3.5-flash
+library_name: transformers
+---
+
+# Model Overview
+
+- **Model Architecture:** Step3p5ForCausalLM
+  - **Input:** Text
+  - **Output:** Text
+- **Supported Hardware Microarchitecture:** AMD MI350/MI355
+- **ROCm**: 7.1.0
+- **PyTorch**: 2.10.0
+- **Transformers**: 4.57.6
+- **Operating System(s):** Linux
+- **Inference Engine:** [vLLM](https://docs.vllm.ai/en/latest/)
+- **Model Optimizer:** [AMD-Quark](https://quark.docs.amd.com/latest/index.html)
+  - **Weight quantization:** MoE-only, OCP MXFP4, Static
+  - **Activation quantization:** MoE-only, OCP MXFP4, Dynamic
+- **Docker Image:** rocm/vllm-dev@sha256:63f1fe04d87376bb173a1e837fba8610ab2dd77039fe7c9b97195f2a89d4d463
+
+
+# Model Quantization
+
+The model was quantized from [stepfun-ai/Step-3.5-Flash](https://huggingface.co/stepfun-ai/Step-3.5-Flash) using [AMD-Quark](https://quark.docs.amd.com/latest/index.html). The weights and activations are both quantized to MXFP4.
+
+
+**Quantization scripts:**
+```
+cd Quark/examples/torch/language_modeling/llm_ptq/
+python3 step3p5_quantize_quark.py --model_dir $MODEL_DIR \
+                          --num_calib_data 128 \
+                          --multi_gpu  \
+                          --trust_remote_code \
+                          --preset mxfp4_moe_only_no_kvcache
+                          --output_dir $output_dir
+```
+For further details or issues, please refer to the AMD-Quark documentation or contact the respective developers.
+
+# Deployment
+### Use with vLLM
+
+This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backend.
+
+## Evaluation
+The model was evaluated on gsm8k benchmarks using the [vLLM](https://docs.vllm.ai/en/latest/) framework.
+
+### Accuracy
+
+<table>
+  <tr>
+   <td><strong>Benchmark</strong>
+   </td>
+   <td><strong>stepfun-ai/Step-3.5-Flash (bf16)</strong>
+   </td>
+   <td><strong>amd/Step-3.5-Flash-MXFP4 (this model)</strong>
+   </td>
+   <td><strong>Recovery</strong>
+   </td>
+  </tr>
+  <tr>
+   <td>gsm8k (flexible-extract) 
+   </td>
+   <td>0.8939
+   </td>
+   <td>0.8726
+   </td>
+   <td>97.6%
+   </td>
+  </tr>
+</table>
+
+
+### Reproduction
+
+The GSM8K results were obtained using the vLLM framework, based on the Docker image `rocm/vllm:nightly`.
+
+#### Note: Due to model support issues in vLLM for Step-3.5-Flash, a few patches need to be applied (specified below) in order to run inference and evaluation using vLLM.  
+
+#### Preparation in container
+```
+# Reinstall vLLM
+pip uninstall vllm -y
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+git checkout de7dd634b969adc6e5f50cff0cc09c1be1711d01
+pip install -r requirements/rocm.txt
+python setup.py develop
+cd ..
+export QUARK_MXFP4_IMPL="triton"
+```
+Modify `vllm/model_executor/models/step3p5.py` by adding the below packed_modules_mapping update in the model's `__init__` function:
+```
+...
+@support_torch_compile
+class Step3p5Model(nn.Module):
+    def __init__(self, vllm_config: VllmConfig, prefix: str = "") -> None:
+        super().__init__()
+
+        self.vllm_config = vllm_config
+        config = vllm_config.model_config.hf_config
++       # Update packed_modules_mapping for quantization
++       if hasattr(vllm_config, "quant_config") and vllm_config.quant_config:
++           vllm_config.quant_config.packed_modules_mapping.update({
++               "qkv_proj": ["q_proj", "k_proj", "v_proj"],
++               "gate_up_proj": ["gate_proj", "up_proj"],
++           })
+        self.vocab_size = config.vocab_size
+        self.config = config
+        ...
+```
+Additionally, modify the same file (`step3p5.py`) by adding the below MoE expert name mapping to the model's `load_weights` function:
+```
+def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+        config = self.config
+        assert config.num_attention_groups > 1, "Only support GQA"
+        
+        ...
+
+        for name, loaded_weight in weights:
+            if name.startswith("model."):
+                local_name = name[len("model.") :]
+                full_name = name
+            else:
+                local_name = name
+                full_name = f"model.{name}" if name else "model"
+
++           # Normalize legacy MoE expert naming like ".moe.<E>.gate_proj" to
++           # the ".moe.experts.<E>.gate_proj" format 
++           if ".moe.experts." not in local_name and ".moe." in local_name:
++               parts = local_name.split(".moe.", 1)
++               if len(parts) == 2 and "." in parts[1]:
++                   expert_and_rest = parts[1]
++                   expert_id, remainder = expert_and_rest.split(".", 1)
++                   if expert_id.isdigit():
++                       local_name = f"{parts[0]}.moe.experts.{expert_id}.{remainder}"
+
+            spec_layer = get_spec_layer_idx_from_weight_name(config, full_name)
+            if spec_layer is not None:
+                continue  # skip spec decode layers for main model
+        ...
+``` 
+Finally, modify `vllm/model_executor/layers/quantization/quark/quark_moe.py` by forcing `self.emulate` to "True":
+```
+class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
+    def __init__(...):
+        super().__init__(moe)
+        ...
+
+        self.model_type = getattr(
+            get_current_vllm_config().model_config.hf_config, "model_type", None
+        )
+
+-       self.emulate = (
+-           not current_platform.supports_mx()
+-           or not self.ocp_mx_scheme.startswith("w_mxfp4")
+-       ) and (self.mxfp4_backend is None or not self.use_rocm_aiter_moe)
++       self.emulate = True
+        
+            logger.warning_once(
+        ...
+```
+
+
+#### Evaluating model using lm_eval
+```
+lm_eval  --model vllm  --model_args 'pretrained=$MODEL_DIR,attention_backend=ROCM_AITER_UNIFIED_ATTN,quantization='quark',trust_remote_code=True'  --tasks gsm8k  --batch_size auto
+```
+
+
+# License
+Modifications Copyright(c) 2026 Advanced Micro Devices, Inc. All rights reserved.

chat_template.jinja

@@ -0,0 +1,80 @@
+{% macro render_content(content) %}{% if content is none %}{{- '' }}{% elif content is string %}{{- content }}{% elif content is mapping %}{{- content['value'] if 'value' in content else content['text'] }}{% elif content is iterable %}{% for item in content %}{% if item.type == 'text' %}{{- item['value'] if 'value' in item else item['text'] }}{% elif item.type == 'image' %}<im_patch>{% endif %}{% endfor %}{% endif %}{% endmacro %}
+{{bos_token}}{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- render_content(messages[0].content) + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions in JSONSchema format:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson(ensure_ascii=False) }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...>\n...\n</function> block must be nested within <tool_call>\n...\n</tool_call> XML tags\n- Required parameters MUST be specified\n</IMPORTANT><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + render_content(messages[0].content) + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and render_content(message.content) is string and not(render_content(message.content).startswith('<tool_response>') and render_content(message.content).endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content) %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {%- set role_name = 'observation' if (message.role == "system" and not loop.first and message.name == 'observation') else message.role %}
+        {{- '<|im_start|>' + role_name + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = render_content(message.reasoning_content) %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- else %}
+                {%- set reasoning_content = '' %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- if tool_call.arguments is defined %}
+                    {%- set arguments = tool_call.arguments %}
+                    {%- for args_name, args_value in arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson(ensure_ascii=False) | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>tool_response\n' }}
+        {%- endif %}
+        {{- '<tool_response>' }}
+        {{- content }}
+        {{- '</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<think>\n' }}
+{%- endif %}

config.json

@@ -0,0 +1,784 @@
+{
+  "architectures": [
+    "Step3p5ForCausalLM"
+  ],
+  "att_impl_type": "GQA",
+  "attention_other_setting": {
+    "attention_type": "sliding_attention",
+    "head_dim": 128,
+    "num_attention_groups": 8,
+    "num_attention_heads": 96,
+    "true_head_dim": 128
+  },
+  "auto_map": {
+    "AutoConfig": "configuration_step3p5.Step3p5Config",
+    "AutoModelForCausalLM": "modeling_step3p5.Step3p5ForCausalLM"
+  },
+  "bos_token_id": 0,
+  "dtype": "bfloat16",
+  "eos_token_id": [
+    1,
+    2,
+    128007
+  ],
+  "head_dim": 128,
+  "hidden_size": 4096,
+  "intermediate_size": 11264,
+  "layer_types": [
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention"
+  ],
+  "max_position_embeddings": 262144,
+  "max_seq_len": 262144,
+  "model_type": "step3p5",
+  "moe_every_n_layer": 1,
+  "moe_intermediate_size": 1280,
+  "moe_layer_offset": 0,
+  "moe_layers_enum": "3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44",
+  "moe_num_experts": 288,
+  "moe_router_activation": "sigmoid",
+  "moe_router_scaling_factor": 3.0,
+  "moe_top_k": 8,
+  "need_fp32_gate": true,
+  "norm_expert_weight": true,
+  "num_attention_groups": 8,
+  "num_attention_heads": 64,
+  "num_hidden_layers": 45,
+  "num_nextn_predict_layers": 3,
+  "partial_rotary_factor": 0.5,
+  "partial_rotary_factors": [
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0,
+    0.5,
+    1.0,
+    1.0,
+    1.0
+  ],
+  "quantization_config": {
+    "algo_config": null,
+    "exclude": [
+      "model.embed_tokens",
+      "model.layers.0.self_attn.q_proj",
+      "model.layers.0.self_attn.k_proj",
+      "model.layers.0.self_attn.v_proj",
+      "model.layers.0.self_attn.o_proj",
+      "model.layers.0.self_attn.g_proj",
+      "model.layers.0.mlp.gate_proj",
+      "model.layers.0.mlp.up_proj",
+      "model.layers.0.mlp.down_proj",
+      "model.layers.1.self_attn.q_proj",
+      "model.layers.1.self_attn.k_proj",
+      "model.layers.1.self_attn.v_proj",
+      "model.layers.1.self_attn.o_proj",
+      "model.layers.1.self_attn.g_proj",
+      "model.layers.1.mlp.gate_proj",
+      "model.layers.1.mlp.up_proj",
+      "model.layers.1.mlp.down_proj",
+      "model.layers.2.self_attn.q_proj",
+      "model.layers.2.self_attn.k_proj",
+      "model.layers.2.self_attn.v_proj",
+      "model.layers.2.self_attn.o_proj",
+      "model.layers.2.self_attn.g_proj",
+      "model.layers.2.mlp.gate_proj",
+      "model.layers.2.mlp.up_proj",
+      "model.layers.2.mlp.down_proj",
+      "model.layers.3.self_attn.q_proj",
+      "model.layers.3.self_attn.k_proj",
+      "model.layers.3.self_attn.v_proj",
+      "model.layers.3.self_attn.o_proj",
+      "model.layers.3.self_attn.g_proj",
+      "model.layers.3.moe.gate",
+      "model.layers.3.share_expert.gate_proj",
+      "model.layers.3.share_expert.up_proj",
+      "model.layers.3.share_expert.down_proj",
+      "model.layers.4.self_attn.q_proj",
+      "model.layers.4.self_attn.k_proj",
+      "model.layers.4.self_attn.v_proj",
+      "model.layers.4.self_attn.o_proj",
+      "model.layers.4.self_attn.g_proj",
+      "model.layers.4.moe.gate",
+      "model.layers.4.share_expert.gate_proj",
+      "model.layers.4.share_expert.up_proj",
+      "model.layers.4.share_expert.down_proj",
+      "model.layers.5.self_attn.q_proj",
+      "model.layers.5.self_attn.k_proj",
+      "model.layers.5.self_attn.v_proj",
+      "model.layers.5.self_attn.o_proj",
+      "model.layers.5.self_attn.g_proj",
+      "model.layers.5.moe.gate",
+      "model.layers.5.share_expert.gate_proj",
+      "model.layers.5.share_expert.up_proj",
+      "model.layers.5.share_expert.down_proj",
+      "model.layers.6.self_attn.q_proj",
+      "model.layers.6.self_attn.k_proj",
+      "model.layers.6.self_attn.v_proj",
+      "model.layers.6.self_attn.o_proj",
+      "model.layers.6.self_attn.g_proj",
+      "model.layers.6.moe.gate",
+      "model.layers.6.share_expert.gate_proj",
+      "model.layers.6.share_expert.up_proj",
+      "model.layers.6.share_expert.down_proj",
+      "model.layers.7.self_attn.q_proj",
+      "model.layers.7.self_attn.k_proj",
+      "model.layers.7.self_attn.v_proj",
+      "model.layers.7.self_attn.o_proj",
+      "model.layers.7.self_attn.g_proj",
+      "model.layers.7.moe.gate",
+      "model.layers.7.share_expert.gate_proj",
+      "model.layers.7.share_expert.up_proj",
+      "model.layers.7.share_expert.down_proj",
+      "model.layers.8.self_attn.q_proj",
+      "model.layers.8.self_attn.k_proj",
+      "model.layers.8.self_attn.v_proj",
+      "model.layers.8.self_attn.o_proj",
+      "model.layers.8.self_attn.g_proj",
+      "model.layers.8.moe.gate",
+      "model.layers.8.share_expert.gate_proj",
+      "model.layers.8.share_expert.up_proj",
+      "model.layers.8.share_expert.down_proj",
+      "model.layers.9.self_attn.q_proj",
+      "model.layers.9.self_attn.k_proj",
+      "model.layers.9.self_attn.v_proj",
+      "model.layers.9.self_attn.o_proj",
+      "model.layers.9.self_attn.g_proj",
+      "model.layers.9.moe.gate",
+      "model.layers.9.share_expert.gate_proj",
+      "model.layers.9.share_expert.up_proj",
+      "model.layers.9.share_expert.down_proj",
+      "model.layers.10.self_attn.q_proj",
+      "model.layers.10.self_attn.k_proj",
+      "model.layers.10.self_attn.v_proj",
+      "model.layers.10.self_attn.o_proj",
+      "model.layers.10.self_attn.g_proj",
+      "model.layers.10.moe.gate",
+      "model.layers.10.share_expert.gate_proj",
+      "model.layers.10.share_expert.up_proj",
+      "model.layers.10.share_expert.down_proj",
+      "model.layers.11.self_attn.q_proj",
+      "model.layers.11.self_attn.k_proj",
+      "model.layers.11.self_attn.v_proj",
+      "model.layers.11.self_attn.o_proj",
+      "model.layers.11.self_attn.g_proj",
+      "model.layers.11.moe.gate",
+      "model.layers.11.share_expert.gate_proj",
+      "model.layers.11.share_expert.up_proj",
+      "model.layers.11.share_expert.down_proj",
+      "model.layers.12.self_attn.q_proj",
+      "model.layers.12.self_attn.k_proj",
+      "model.layers.12.self_attn.v_proj",
+      "model.layers.12.self_attn.o_proj",
+      "model.layers.12.self_attn.g_proj",
+      "model.layers.12.moe.gate",
+      "model.layers.12.share_expert.gate_proj",
+      "model.layers.12.share_expert.up_proj",
+      "model.layers.12.share_expert.down_proj",
+      "model.layers.13.self_attn.q_proj",
+      "model.layers.13.self_attn.k_proj",
+      "model.layers.13.self_attn.v_proj",
+      "model.layers.13.self_attn.o_proj",
+      "model.layers.13.self_attn.g_proj",
+      "model.layers.13.moe.gate",
+      "model.layers.13.share_expert.gate_proj",
+      "model.layers.13.share_expert.up_proj",
+      "model.layers.13.share_expert.down_proj",
+      "model.layers.14.self_attn.q_proj",
+      "model.layers.14.self_attn.k_proj",
+      "model.layers.14.self_attn.v_proj",
+      "model.layers.14.self_attn.o_proj",
+      "model.layers.14.self_attn.g_proj",
+      "model.layers.14.moe.gate",
+      "model.layers.14.share_expert.gate_proj",
+      "model.layers.14.share_expert.up_proj",
+      "model.layers.14.share_expert.down_proj",
+      "model.layers.15.self_attn.q_proj",
+      "model.layers.15.self_attn.k_proj",
+      "model.layers.15.self_attn.v_proj",
+      "model.layers.15.self_attn.o_proj",
+      "model.layers.15.self_attn.g_proj",
+      "model.layers.15.moe.gate",
+      "model.layers.15.share_expert.gate_proj",
+      "model.layers.15.share_expert.up_proj",
+      "model.layers.15.share_expert.down_proj",
+      "model.layers.16.self_attn.q_proj",
+      "model.layers.16.self_attn.k_proj",
+      "model.layers.16.self_attn.v_proj",
+      "model.layers.16.self_attn.o_proj",
+      "model.layers.16.self_attn.g_proj",
+      "model.layers.16.moe.gate",
+      "model.layers.16.share_expert.gate_proj",
+      "model.layers.16.share_expert.up_proj",
+      "model.layers.16.share_expert.down_proj",
+      "model.layers.17.self_attn.q_proj",
+      "model.layers.17.self_attn.k_proj",
+      "model.layers.17.self_attn.v_proj",
+      "model.layers.17.self_attn.o_proj",
+      "model.layers.17.self_attn.g_proj",
+      "model.layers.17.moe.gate",
+      "model.layers.17.share_expert.gate_proj",
+      "model.layers.17.share_expert.up_proj",
+      "model.layers.17.share_expert.down_proj",
+      "model.layers.18.self_attn.q_proj",
+      "model.layers.18.self_attn.k_proj",
+      "model.layers.18.self_attn.v_proj",
+      "model.layers.18.self_attn.o_proj",
+      "model.layers.18.self_attn.g_proj",
+      "model.layers.18.moe.gate",
+      "model.layers.18.share_expert.gate_proj",
+      "model.layers.18.share_expert.up_proj",
+      "model.layers.18.share_expert.down_proj",
+      "model.layers.19.self_attn.q_proj",
+      "model.layers.19.self_attn.k_proj",
+      "model.layers.19.self_attn.v_proj",
+      "model.layers.19.self_attn.o_proj",
+      "model.layers.19.self_attn.g_proj",
+      "model.layers.19.moe.gate",
+      "model.layers.19.share_expert.gate_proj",
+      "model.layers.19.share_expert.up_proj",
+      "model.layers.19.share_expert.down_proj",
+      "model.layers.20.self_attn.q_proj",
+      "model.layers.20.self_attn.k_proj",
+      "model.layers.20.self_attn.v_proj",
+      "model.layers.20.self_attn.o_proj",
+      "model.layers.20.self_attn.g_proj",
+      "model.layers.20.moe.gate",
+      "model.layers.20.share_expert.gate_proj",
+      "model.layers.20.share_expert.up_proj",
+      "model.layers.20.share_expert.down_proj",
+      "model.layers.21.self_attn.q_proj",
+      "model.layers.21.self_attn.k_proj",
+      "model.layers.21.self_attn.v_proj",
+      "model.layers.21.self_attn.o_proj",
+      "model.layers.21.self_attn.g_proj",
+      "model.layers.21.moe.gate",
+      "model.layers.21.share_expert.gate_proj",
+      "model.layers.21.share_expert.up_proj",
+      "model.layers.21.share_expert.down_proj",
+      "model.layers.22.self_attn.q_proj",
+      "model.layers.22.self_attn.k_proj",
+      "model.layers.22.self_attn.v_proj",
+      "model.layers.22.self_attn.o_proj",
+      "model.layers.22.self_attn.g_proj",
+      "model.layers.22.moe.gate",
+      "model.layers.22.share_expert.gate_proj",
+      "model.layers.22.share_expert.up_proj",
+      "model.layers.22.share_expert.down_proj",
+      "model.layers.23.self_attn.q_proj",
+      "model.layers.23.self_attn.k_proj",
+      "model.layers.23.self_attn.v_proj",
+      "model.layers.23.self_attn.o_proj",
+      "model.layers.23.self_attn.g_proj",
+      "model.layers.23.moe.gate",
+      "model.layers.23.share_expert.gate_proj",
+      "model.layers.23.share_expert.up_proj",
+      "model.layers.23.share_expert.down_proj",
+      "model.layers.24.self_attn.q_proj",
+      "model.layers.24.self_attn.k_proj",
+      "model.layers.24.self_attn.v_proj",
+      "model.layers.24.self_attn.o_proj",
+      "model.layers.24.self_attn.g_proj",
+      "model.layers.24.moe.gate",
+      "model.layers.24.share_expert.gate_proj",
+      "model.layers.24.share_expert.up_proj",
+      "model.layers.24.share_expert.down_proj",
+      "model.layers.25.self_attn.q_proj",
+      "model.layers.25.self_attn.k_proj",
+      "model.layers.25.self_attn.v_proj",
+      "model.layers.25.self_attn.o_proj",
+      "model.layers.25.self_attn.g_proj",
+      "model.layers.25.moe.gate",
+      "model.layers.25.share_expert.gate_proj",
+      "model.layers.25.share_expert.up_proj",
+      "model.layers.25.share_expert.down_proj",
+      "model.layers.26.self_attn.q_proj",
+      "model.layers.26.self_attn.k_proj",
+      "model.layers.26.self_attn.v_proj",
+      "model.layers.26.self_attn.o_proj",
+      "model.layers.26.self_attn.g_proj",
+      "model.layers.26.moe.gate",
+      "model.layers.26.share_expert.gate_proj",
+      "model.layers.26.share_expert.up_proj",
+      "model.layers.26.share_expert.down_proj",
+      "model.layers.27.self_attn.q_proj",
+      "model.layers.27.self_attn.k_proj",
+      "model.layers.27.self_attn.v_proj",
+      "model.layers.27.self_attn.o_proj",
+      "model.layers.27.self_attn.g_proj",
+      "model.layers.27.moe.gate",
+      "model.layers.27.share_expert.gate_proj",
+      "model.layers.27.share_expert.up_proj",
+      "model.layers.27.share_expert.down_proj",
+      "model.layers.28.self_attn.q_proj",
+      "model.layers.28.self_attn.k_proj",
+      "model.layers.28.self_attn.v_proj",
+      "model.layers.28.self_attn.o_proj",
+      "model.layers.28.self_attn.g_proj",
+      "model.layers.28.moe.gate",
+      "model.layers.28.share_expert.gate_proj",
+      "model.layers.28.share_expert.up_proj",
+      "model.layers.28.share_expert.down_proj",
+      "model.layers.29.self_attn.q_proj",
+      "model.layers.29.self_attn.k_proj",
+      "model.layers.29.self_attn.v_proj",
+      "model.layers.29.self_attn.o_proj",
+      "model.layers.29.self_attn.g_proj",
+      "model.layers.29.moe.gate",
+      "model.layers.29.share_expert.gate_proj",
+      "model.layers.29.share_expert.up_proj",
+      "model.layers.29.share_expert.down_proj",
+      "model.layers.30.self_attn.q_proj",
+      "model.layers.30.self_attn.k_proj",
+      "model.layers.30.self_attn.v_proj",
+      "model.layers.30.self_attn.o_proj",
+      "model.layers.30.self_attn.g_proj",
+      "model.layers.30.moe.gate",
+      "model.layers.30.share_expert.gate_proj",
+      "model.layers.30.share_expert.up_proj",
+      "model.layers.30.share_expert.down_proj",
+      "model.layers.31.self_attn.q_proj",
+      "model.layers.31.self_attn.k_proj",
+      "model.layers.31.self_attn.v_proj",
+      "model.layers.31.self_attn.o_proj",
+      "model.layers.31.self_attn.g_proj",
+      "model.layers.31.moe.gate",
+      "model.layers.31.share_expert.gate_proj",
+      "model.layers.31.share_expert.up_proj",
+      "model.layers.31.share_expert.down_proj",
+      "model.layers.32.self_attn.q_proj",
+      "model.layers.32.self_attn.k_proj",
+      "model.layers.32.self_attn.v_proj",
+      "model.layers.32.self_attn.o_proj",
+      "model.layers.32.self_attn.g_proj",
+      "model.layers.32.moe.gate",
+      "model.layers.32.share_expert.gate_proj",
+      "model.layers.32.share_expert.up_proj",
+      "model.layers.32.share_expert.down_proj",
+      "model.layers.33.self_attn.q_proj",
+      "model.layers.33.self_attn.k_proj",
+      "model.layers.33.self_attn.v_proj",
+      "model.layers.33.self_attn.o_proj",
+      "model.layers.33.self_attn.g_proj",
+      "model.layers.33.moe.gate",
+      "model.layers.33.share_expert.gate_proj",
+      "model.layers.33.share_expert.up_proj",
+      "model.layers.33.share_expert.down_proj",
+      "model.layers.34.self_attn.q_proj",
+      "model.layers.34.self_attn.k_proj",
+      "model.layers.34.self_attn.v_proj",
+      "model.layers.34.self_attn.o_proj",
+      "model.layers.34.self_attn.g_proj",
+      "model.layers.34.moe.gate",
+      "model.layers.34.share_expert.gate_proj",
+      "model.layers.34.share_expert.up_proj",
+      "model.layers.34.share_expert.down_proj",
+      "model.layers.35.self_attn.q_proj",
+      "model.layers.35.self_attn.k_proj",
+      "model.layers.35.self_attn.v_proj",
+      "model.layers.35.self_attn.o_proj",
+      "model.layers.35.self_attn.g_proj",
+      "model.layers.35.moe.gate",
+      "model.layers.35.share_expert.gate_proj",
+      "model.layers.35.share_expert.up_proj",
+      "model.layers.35.share_expert.down_proj",
+      "model.layers.36.self_attn.q_proj",
+      "model.layers.36.self_attn.k_proj",
+      "model.layers.36.self_attn.v_proj",
+      "model.layers.36.self_attn.o_proj",
+      "model.layers.36.self_attn.g_proj",
+      "model.layers.36.moe.gate",
+      "model.layers.36.share_expert.gate_proj",
+      "model.layers.36.share_expert.up_proj",
+      "model.layers.36.share_expert.down_proj",
+      "model.layers.37.self_attn.q_proj",
+      "model.layers.37.self_attn.k_proj",
+      "model.layers.37.self_attn.v_proj",
+      "model.layers.37.self_attn.o_proj",
+      "model.layers.37.self_attn.g_proj",
+      "model.layers.37.moe.gate",
+      "model.layers.37.share_expert.gate_proj",
+      "model.layers.37.share_expert.up_proj",
+      "model.layers.37.share_expert.down_proj",
+      "model.layers.38.self_attn.q_proj",
+      "model.layers.38.self_attn.k_proj",
+      "model.layers.38.self_attn.v_proj",
+      "model.layers.38.self_attn.o_proj",
+      "model.layers.38.self_attn.g_proj",
+      "model.layers.38.moe.gate",
+      "model.layers.38.share_expert.gate_proj",
+      "model.layers.38.share_expert.up_proj",
+      "model.layers.38.share_expert.down_proj",
+      "model.layers.39.self_attn.q_proj",
+      "model.layers.39.self_attn.k_proj",
+      "model.layers.39.self_attn.v_proj",
+      "model.layers.39.self_attn.o_proj",
+      "model.layers.39.self_attn.g_proj",
+      "model.layers.39.moe.gate",
+      "model.layers.39.share_expert.gate_proj",
+      "model.layers.39.share_expert.up_proj",
+      "model.layers.39.share_expert.down_proj",
+      "model.layers.40.self_attn.q_proj",
+      "model.layers.40.self_attn.k_proj",
+      "model.layers.40.self_attn.v_proj",
+      "model.layers.40.self_attn.o_proj",
+      "model.layers.40.self_attn.g_proj",
+      "model.layers.40.moe.gate",
+      "model.layers.40.share_expert.gate_proj",
+      "model.layers.40.share_expert.up_proj",
+      "model.layers.40.share_expert.down_proj",
+      "model.layers.41.self_attn.q_proj",
+      "model.layers.41.self_attn.k_proj",
+      "model.layers.41.self_attn.v_proj",
+      "model.layers.41.self_attn.o_proj",
+      "model.layers.41.self_attn.g_proj",
+      "model.layers.41.moe.gate",
+      "model.layers.41.share_expert.gate_proj",
+      "model.layers.41.share_expert.up_proj",
+      "model.layers.41.share_expert.down_proj",
+      "model.layers.42.self_attn.q_proj",
+      "model.layers.42.self_attn.k_proj",
+      "model.layers.42.self_attn.v_proj",
+      "model.layers.42.self_attn.o_proj",
+      "model.layers.42.self_attn.g_proj",
+      "model.layers.42.moe.gate",
+      "model.layers.42.share_expert.gate_proj",
+      "model.layers.42.share_expert.up_proj",
+      "model.layers.42.share_expert.down_proj",
+      "model.layers.43.self_attn.q_proj",
+      "model.layers.43.self_attn.k_proj",
+      "model.layers.43.self_attn.v_proj",
+      "model.layers.43.self_attn.o_proj",
+      "model.layers.43.self_attn.g_proj",
+      "model.layers.43.moe.gate",
+      "model.layers.43.share_expert.gate_proj",
+      "model.layers.43.share_expert.up_proj",
+      "model.layers.43.share_expert.down_proj",
+      "model.layers.44.self_attn.q_proj",
+      "model.layers.44.self_attn.k_proj",
+      "model.layers.44.self_attn.v_proj",
+      "model.layers.44.self_attn.o_proj",
+      "model.layers.44.self_attn.g_proj",
+      "model.layers.44.moe.gate",
+      "model.layers.44.share_expert.gate_proj",
+      "model.layers.44.share_expert.up_proj",
+      "model.layers.44.share_expert.down_proj",
+      "lm_head"
+    ],
+    "export": {
+      "kv_cache_group": [],
+      "min_kv_scale": 0.0,
+      "pack_method": "reorder",
+      "weight_format": "real_quantized",
+      "weight_merge_groups": null
+    },
+    "global_quant_config": {
+      "bias": null,
+      "input_tensors": {
+        "ch_axis": -1,
+        "dtype": "fp4",
+        "group_size": 32,
+        "is_dynamic": true,
+        "is_scale_quant": false,
+        "mx_element_dtype": null,
+        "observer_cls": "PerBlockMXObserver",
+        "qscheme": "per_group",
+        "round_method": "half_even",
+        "scale_calculation_mode": "even",
+        "scale_format": "e8m0",
+        "scale_type": "float",
+        "symmetric": null
+      },
+      "output_tensors": null,
+      "target_device": null,
+      "weight": {
+        "ch_axis": -1,
+        "dtype": "fp4",
+        "group_size": 32,
+        "is_dynamic": false,
+        "is_scale_quant": false,
+        "mx_element_dtype": null,
+        "observer_cls": "PerBlockMXObserver",
+        "qscheme": "per_group",
+        "round_method": "half_even",
+        "scale_calculation_mode": "even",
+        "scale_format": "e8m0",
+        "scale_type": "float",
+        "symmetric": null
+      }
+    },
+    "kv_cache_post_rope": false,
+    "kv_cache_quant_config": {},
+    "layer_quant_config": {},
+    "layer_type_quant_config": {},
+    "quant_method": "quark",
+    "quant_mode": "eager_mode",
+    "softmax_quant_spec": null,
+    "version": "0.12+422c9a6d36"
+  },
+  "rms_norm_eps": 1e-05,
+  "rope_parameters": {
+    "factor": 2.0,
+    "high_freq_factor": 32.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 131072,
+    "rope_type": "llama3"
+  },
+  "rope_scaling": {
+    "factor": 2.0,
+    "high_freq_factor": 32.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 131072,
+    "rope_type": "llama3"
+  },
+  "rope_theta": [
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0,
+    5000000.0,
+    10000.0,
+    10000.0,
+    10000.0
+  ],
+  "share_expert_dim": 1280,
+  "sink": false,
+  "sliding_window": 512,
+  "swiglu_limits": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    7,
+    7,
+    0.0,
+    0.0,
+    0.0
+  ],
+  "swiglu_limits_shared": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    16,
+    0.0,
+    0.0,
+    0.0
+  ],
+  "transformers_version": "4.57.6",
+  "use_cache": false,
+  "use_head_wise_attn_gate": true,
+  "use_moe": true,
+  "use_moe_router_bias": true,
+  "use_qk_norm": true,
+  "use_rope_layers": [],
+  "vocab_size": 128896,
+  "yarn_only_types": [
+    "full_attention"
+  ],
+  "zero_centered": true
+}

configuration_step3p5.py

@@ -0,0 +1,59 @@
+from typing import Any, Optional, Union
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+
+class Step3p5Config(PretrainedConfig):
+    model_type = "step3p5"
+    architectures = ["Step3p5ForCausalLM"]
+
+    def __init__(
+        self,
+        hidden_size: int = 4096,
+        intermediate_size: int = 11264,
+        num_attention_heads: int = 64,
+        num_attention_groups: int = 8,
+        num_hidden_layers: int = 45,
+        max_seq_len: int = 128000,
+        vocab_size: int = 128815,
+        rms_norm_eps: float = 1e-5,
+        moe_intermediate_size: int = 1280,
+        moe_num_experts: int = 288,
+        moe_top_k: int = 8,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[dict[str, Any]] = None,
+        max_position_embeddings: int = 128000,
+        share_expert_dims: int = 1280,
+        head_dim: int = 128,
+        norm_expert_weight: bool = True,
+        layer_types: list[str] = None,
+        sliding_window: Optional[int] = None,
+        moe_layers_enum: tuple[int] = (3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
+                                       15, 16, 17, 18, 19, 20, 21, 22, 23, 24,
+                                       25, 26, 27, 28, 29, 30, 31, 32, 33, 34,
+                                       35, 36, 37, 38, 39, 40, 41, 42, 43, 44),
+        **kwargs,
+    ) -> None:
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.num_attention_groups = num_attention_groups
+        self.num_hidden_layers = num_hidden_layers
+        self.max_seq_len = max_seq_len
+        self.vocab_size = vocab_size
+        self.rms_norm_eps = rms_norm_eps
+        self.moe_intermediate_size = moe_intermediate_size
+        self.moe_num_experts = moe_num_experts
+        self.moe_top_k = moe_top_k
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.max_position_embeddings = max_position_embeddings
+        self.share_expert_dim = share_expert_dims
+        self.head_dim = head_dim
+        self.norm_expert_weight = norm_expert_weight
+        self.moe_layers_enum = moe_layers_enum
+        self.layer_types = layer_types
+        self.sliding_window = sliding_window
+        super().__init__(**kwargs)
+

generation_config.json

@@ -0,0 +1,11 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 0,
+  "eos_token_id": [
+    1,
+    2,
+    128007
+  ],
+  "transformers_version": "4.57.6",
+  "use_cache": false
+}

model-00001-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:384546f5d09641de80e4878241ab44da14cbd90c21fa9995c334a9888bea5272
+size 4997829840

model-00002-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed5cea6a5f458a8cc3c59c08403ca350927edc2f206c05e482e32f678eec0de5
+size 4997967800

model-00003-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bef1ec40bbb87937f8777f6e57bd965408395f923bd9512ebd528ffbd0d99c44
+size 4998486552

model-00004-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe03c3c2c53f90afa5de4c5cb905115a4556e2ac9fba2a39e3d9cb1e1e1245fb
+size 4997967808

model-00005-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54775d7e4e6eb6ea1f6ac958ac9be3ff05e07aa727ba855af9eb235629c15af8
+size 4998489136

model-00006-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a30e99ef04b3e91797e05986317852fd132658eab8650c34799f2d5ec1f3d6d1
+size 4997971104

model-00007-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1622bc5c0bc51eeda8b4fe4a2316dcffb1c57febaf43f88a4ab43c14b7d091f4
+size 4998489992

model-00008-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33e700b1470e7d3393aabb30d0a3786ae6b13920d9efcdee62e7a15c93141691
+size 4997971288

model-00009-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7206bd64c85d4eaadea5bb67c988494042bf6a86a86702a913d8e36d93e62ca4
+size 4998490088

model-00010-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:104a1aef751338400e8684e1283d223837b6993f847735d4b7f2656fa9052636
+size 4993155312

model-00011-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29ca0e3dbd4b8b5100967845407362b203e26950a685790b7e5598613b77b19a
+size 4998489800

model-00012-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e447996b9b05e4cefffe76ca8c0c30cc8b297fe14e5a8525981184f75f35f32
+size 4997971104

model-00013-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0410ad3ca30cab2f3a98d5fbb0a82fd9fe55224ff7acc4706019faa2eea4da3
+size 4998489800

model-00014-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ff3a1a570903c6681e62a6110b31cbd6c590e3b436ca50dd7bc367e0fc65986
+size 4997971104

model-00015-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:402aa7a977a0ff3c79d718fc5077b29cd9a0d56176fc96f8ac401e89677a028f
+size 4998489808

model-00016-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06ca8045ca30b180ffb580ad74a33f3eb9b8fde6c18936761bab6f669bfad400
+size 4997971208

model-00017-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd45f815bf5b5a30ed5ce73850b18f21c00e39c266befb63bc4b1e9ebb43d424
+size 4998490032

model-00018-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3894225b6a72e4c30b14571d88934cbd9f91f70652347311324f9c557d1d8ed7
+size 4997971288

model-00019-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4609212a52dc17a26b1ffaf76e8d6b4d2b13ad07e57cf6a5ab8e39e18c052d4b
+size 4986151200

model-00020-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2be2764df435b6a37e7a4207f4a881d82604e6e032587f958f8a84f5fec5a11f
+size 5000277752

model-00021-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c83bd8043b4210f5b0ed7a6ec638958a253799c04e39eaea5afcc6f30803b9a
+size 4998135064

model-00022-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d136448b1e20d79445f1d531a4c4ecc229cbdfc4f3f2d368783054dfcd43a20f
+size 4998489800

model-00023-of-00023.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec8c5e201893b3061feb026228557356348afef0aa09765c679a81961d571e8f
+size 4540164960

modeling_step3p5.py

@@ -0,0 +1,900 @@
+# Copyright 2025 The LLAMA4 and HuggingFace Inc. team. All rights reserved.
+#
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Callable, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.generation import GenerationMixin
+from transformers.masking_utils import (create_causal_mask,
+                                        create_sliding_window_causal_mask)
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import BaseModelOutputWithPast, ModelOutput
+from transformers.modeling_rope_utils import (ROPE_INIT_FUNCTIONS,
+                                              dynamic_rope_update)
+from transformers.modeling_utils import (ALL_ATTENTION_FUNCTIONS,
+                                         PreTrainedModel)
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs, can_return_tuple, logging
+
+from .configuration_step3p5 import Step3p5Config
+
+logger = logging.get_logger(__name__)
+
+__all__ = ["Step3p5Model", "Step3p5ForCausalLM"]
+
+class Step3p5RotaryEmbedding(nn.Module):
+
+    def __init__(self, config: Step3p5Config, device=None, layer_idx=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        self.layer_idx = layer_idx
+        if config.rope_parameters is not None:
+            self.rope_type = config.rope_parameters.get(
+                "rope_type", config.rope_parameters.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        partial_rotary_factors = getattr(config, "partial_rotary_factors",
+                                         None)
+        if partial_rotary_factors is not None:
+            config.partial_rotary_factor = partial_rotary_factors[
+                self.layer_idx]
+        else:
+            config.partial_rotary_factor = 1.0
+
+        self.rope_theta = config.rope_theta
+        if isinstance(config.rope_theta, list):
+            self.rope_theta = config.rope_theta.copy()
+            config.rope_theta = self.rope_theta[self.layer_idx]
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = self.rope_init_fn(
+            self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+        config.rope_theta = self.rope_theta
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(
+            position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float().to(x.device)
+
+        device_type = x.device.type if isinstance(
+            x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type,
+                            enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float()
+                     @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., :x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    rotary_dim = cos.shape[-1]
+    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
+    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]
+
+    # Apply rotary embeddings on the first half or full tensor
+    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
+    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)
+
+    # Concatenate back to full shape
+    q_embed = torch.cat([q_embed, q_pass], dim=-1)
+    k_embed = torch.cat([k_embed, k_pass], dim=-1)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :,
+                                  None, :, :].expand(batch,
+                                                     num_key_value_heads,
+                                                     n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen,
+                                 head_dim)
+
+
+# Adapted from transformers.models.llama.modeling_llama.eager_attention_forward -> llama4 doesn't cast attn weights to fp32
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+    # breakpoint()
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, :key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+    attn_weights = nn.functional.dropout(attn_weights,
+                                         p=dropout,
+                                         training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+@dataclass
+class Step3p5CausalLMOutputWithPast(ModelOutput):
+    r"""
+    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+        Language modeling loss (for next-token prediction).
+    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+        Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+        `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+        `past_key_values` input) to speed up sequential decoding.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[list[torch.FloatTensor]] = None
+    hidden_states: Optional[tuple[torch.FloatTensor]] = None
+    attentions: Optional[tuple[torch.FloatTensor]] = None
+
+
+class Step3p5MLP(nn.Module):
+
+    def __init__(self, config, intermediate_size=None, swiglu_limit=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size if intermediate_size is not None else config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size,
+                                   self.intermediate_size,
+                                   bias=False)
+        self.up_proj = nn.Linear(self.hidden_size,
+                                 self.intermediate_size,
+                                 bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size,
+                                   self.hidden_size,
+                                   bias=False)
+        self.act_fn = ACT2FN["silu"]
+        self.limit = swiglu_limit
+
+    def forward(self, x):
+        up = self.up_proj(x)
+        gate = self.act_fn(self.gate_proj(x))
+        if self.limit is not None:
+            gate = gate.clamp(min=None, max=self.limit)
+            up = up.clamp(min=-self.limit, max=self.limit)
+
+        return self.down_proj(gate * up)
+
+
+def sigmoid_routing_function(gating_output: torch.Tensor, topk: int,
+                             renormalize: bool):
+    gating_output = gating_output.float()
+    gate_prob = torch.sigmoid(gating_output)
+    gate_prob = gate_prob / gate_prob.sum(dim=-1, keepdim=True)
+    topk_prob, indices = torch.topk(gate_prob, k=topk, dim=1)
+    expert_topk_weight = topk_prob
+    if renormalize:
+        expert_topk_weight = expert_topk_weight / torch.sum(
+            expert_topk_weight, dim=-1, keepdim=True)
+    return expert_topk_weight, indices
+
+
+def softmax_routing_function(gating_output: torch.Tensor, top_k: int,
+                             renormalize: bool):
+    gating_output = gating_output.float()
+    gate_prob = torch.softmax(gating_output, dim=-1)
+    gate_prob = gate_prob / gate_prob.sum(dim=-1, keepdim=True)
+    topk_prob, indices = torch.topk(gate_prob, k=top_k, dim=1)
+    expert_topk_weight = topk_prob
+    if renormalize:
+        expert_topk_weight = expert_topk_weight / torch.sum(
+            expert_topk_weight, dim=-1, keepdim=True)
+    return expert_topk_weight, indices.to(torch.int32)
+
+
+class MoELinear(nn.Module):
+
+    def __init__(self, num_experts, in_features, out_features):
+        super().__init__()
+        self.num_experts = num_experts
+        self.in_features = in_features
+        self.out_features = out_features
+        self.weight = nn.Parameter(
+            torch.empty(num_experts, out_features, in_features))
+
+    def forward(self, x, expert_id):
+        x = F.linear(x.float(), self.weight[expert_id].float())
+        return x
+
+
+class Step3p5MoEMLP(nn.Module):
+
+    def __init__(self, config, swiglu_limit=None):
+        super().__init__()
+        self.num_experts = config.moe_num_experts
+        self.top_k = config.moe_top_k
+        self.hidden_size = config.hidden_size
+        self.moe_intermediate_size = config.moe_intermediate_size
+
+        self.use_moe_router_bias = config.use_moe_router_bias
+        if self.use_moe_router_bias:
+            self.router_bias = nn.Parameter(torch.zeros(config.moe_num_experts,
+                                                        dtype=torch.float32),
+                                            requires_grad=False)
+            self.custom_routing_function = self.router_bias_func
+        elif config.moe_router_activation == "sigmoid":
+            self.custom_routing_function = sigmoid_routing_function
+        else:
+            self.custom_routing_function = None
+        self.need_fp32_gate = config.need_fp32_gate
+        self.routed_scaling_factor = getattr(config,
+                                             "moe_router_scaling_factor", 1.0)
+        
+        # gating
+        self.gate = nn.Linear(self.hidden_size, self.num_experts, bias=False)
+            
+        self.act_fn = ACT2FN["silu"]
+        self.limit = swiglu_limit
+
+        self.up_proj = MoELinear(self.num_experts, self.hidden_size,
+                                 self.moe_intermediate_size)
+        self.gate_proj = MoELinear(self.num_experts, self.hidden_size,
+                                   self.moe_intermediate_size)
+        self.down_proj = MoELinear(self.num_experts,
+                                   self.moe_intermediate_size,
+                                   self.hidden_size)
+
+    def router_bias_func(self, gating_output: torch.Tensor, topk: int,
+                         renormalize: bool):
+        gate_prob = torch.sigmoid(gating_output.float())
+        gate_prob_with_bias = gate_prob + self.router_bias.unsqueeze(0)
+        _, indices = torch.topk(gate_prob_with_bias, k=topk, dim=1)
+        topk_prob = torch.gather(gate_prob, 1, indices)
+        expert_topk_weight = topk_prob
+        if renormalize:
+            expert_topk_weight = expert_topk_weight / (
+                torch.sum(expert_topk_weight, dim=-1, keepdim=True) + 1e-20)
+        return expert_topk_weight, indices
+
+    def get_expert_output(self, inputs: torch.Tensor, expert_id):
+        #if self.limit is None:
+        up = self.up_proj(inputs, expert_id)
+        gate = self.act_fn(self.gate_proj(inputs, expert_id))
+        if self.limit is not None:
+            gate = gate.clamp(min=None, max=self.limit)
+            up = up.clamp(min=-self.limit, max=self.limit)
+
+        return self.down_proj(gate * up, expert_id)
+
+    def forward(self, hidden_states):
+        """ """
+        batch_size, sequence_length, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        if self.need_fp32_gate:
+            router_logits = torch.matmul(hidden_states.to(torch.float32), self.gate.weight.t().to(torch.float32))
+        else:
+            # router_logits: (batch * sequence_length, n_experts)
+            router_logits = self.gate(hidden_states)
+        
+        if self.custom_routing_function:
+            routing_weights, selected_experts = self.custom_routing_function(
+                router_logits, self.top_k, renormalize=True)
+        else:
+            routing_weights = F.softmax(router_logits,
+                                        dim=1,
+                                        dtype=torch.float)
+            routing_weights, selected_experts = torch.topk(routing_weights,
+                                                           self.top_k,
+                                                           dim=-1)
+
+        routing_weights = routing_weights * self.routed_scaling_factor
+
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim),
+            dtype=hidden_states.dtype,
+            device=hidden_states.device)
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(
+            selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = (
+                self.get_expert_output(current_state, expert_idx) *
+                routing_weights[top_x, idx, None])
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(
+                0, top_x, current_hidden_states.to(hidden_states.dtype))
+        final_hidden_states = final_hidden_states.reshape(
+            batch_size, sequence_length, hidden_dim)
+        return final_hidden_states
+
+
+class Step3p5RMSNorm(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        eps: float = 1e-5,
+    ) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        dtype = x.dtype
+        x = x.float()
+        variance = x.pow(2).mean(dim=-1, keepdim=True)
+        normed = x * torch.rsqrt(variance + self.variance_epsilon)
+        normed = normed * (self.weight.float() + 1)
+        return normed.to(dtype)
+class Step3p5Attention(nn.Module):
+
+    def __init__(self, config: Step3p5Config, layer_idx):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.num_attention_heads = config.num_attention_heads
+        self.num_key_value_heads = config.num_attention_groups
+
+        layer_types = getattr(config, "layer_types", [])
+        if layer_types:
+            enable_sliding_window = layer_types[
+                self.layer_idx] == "sliding_attention"
+        else:
+            enable_sliding_window = self.layer_idx % 2 == 0
+        
+        if hasattr(config, "yarn_only_types") and layer_types[
+                self.layer_idx] not in config.yarn_only_types:
+            config.rope_parameters = None
+        else:
+            config.rope_parameters = getattr(config, "rope_scaling", None)
+
+        self.sliding_window = config.sliding_window
+        if enable_sliding_window:
+            self.num_attention_heads = config.attention_other_setting[
+                "num_attention_heads"]
+            self.num_key_value_heads = config.attention_other_setting[
+                "num_attention_groups"]
+
+        if self.sliding_window is not None and enable_sliding_window:
+            self.sliding_window = (self.sliding_window)
+        else:
+            self.sliding_window = None
+        self.head_dim = getattr(config, "head_dim",
+                        config.hidden_size // self.num_attention_heads)
+        self.num_key_value_groups = self.num_attention_heads // self.num_key_value_heads
+
+        self.rotary_emb = Step3p5RotaryEmbedding(config, layer_idx=layer_idx)
+
+        self.q_size = self.num_attention_heads * self.head_dim
+        self.kv_size = self.num_key_value_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(config.hidden_size, self.q_size, bias=False)
+        self.k_proj = nn.Linear(config.hidden_size, self.kv_size, bias=False)
+        self.v_proj = nn.Linear(config.hidden_size, self.kv_size, bias=False)
+        self.o_proj = nn.Linear(self.q_size, config.hidden_size, bias=False)
+        self.q_norm = Step3p5RMSNorm(self.head_dim,
+                                    eps=config.rms_norm_eps)
+        self.k_norm = Step3p5RMSNorm(self.head_dim,
+                                    eps=config.rms_norm_eps)
+
+        self.use_head_wise_attn_gate = config.use_head_wise_attn_gate
+        if self.use_head_wise_attn_gate:
+            self.g_proj = nn.Linear(config.hidden_size,
+                                    self.num_attention_heads,
+                                    bias=False)
+
+        self.use_rope = True
+        use_rope_layers = getattr(config, "use_rope_layers", None)
+        if use_rope_layers:
+            self.use_rope = use_rope_layers[self.layer_idx]
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor],
+               Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_norm(
+            self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        key_states = self.k_norm(
+            self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(
+            1, 2)
+        if self.use_head_wise_attn_gate:
+            gate_states = self.g_proj(hidden_states)
+        cos, sin = self.rotary_emb(hidden_states, position_ids)
+
+        # cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin)
+
+        # query_states, key_states = apply_rotary_pos_emb(query_norm_states, key_norm_states, cos, sin)
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; position_ids needed for the static cache
+            cache_kwargs = {
+                "sin": sin,
+                "cos": cos,
+                "cache_position": cache_position
+            }
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        # TODO: considering FP8；
+        # RuntimeError: Expected attn_mask dtype to be bool or float or to match query dtype,
+        # but got attn_mask.dtype: long int and  query.dtype: c10::BFloat16 instead.
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[
+                self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sliding_window=self.sliding_window,  # main diff with Llama
+            **kwargs,
+        )
+        attn_output = attn_output.reshape(*input_shape, -1)
+        if self.use_head_wise_attn_gate:
+            output = attn_output.view(
+                *attn_output.shape[:-1], self.num_attention_heads,
+                self.head_dim) * gate_states.unsqueeze(-1).sigmoid()
+            attn_output = output.view(*attn_output.shape)
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class Step3p5DecoderLayer(GradientCheckpointingLayer):
+
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.layer_idx = layer_idx
+        self.self_attn = Step3p5Attention(config, layer_idx)
+        self.attention_type = config.layer_types[layer_idx]
+
+        moe_layers_enum = getattr(config, "moe_layers_enum", None)
+        if moe_layers_enum is not None:
+            moe_layers_idx = [
+                int(i) for i in moe_layers_enum.strip().split(',')
+            ]
+        else:
+            moe_layers_idx = [i for i in range(1, config.num_hidden_layers)]
+        self.is_moe_layer = layer_idx in moe_layers_idx
+        self.use_moe = False
+
+        if config.swiglu_limits_shared and config.swiglu_limits_shared[
+                layer_idx] is not None and config.swiglu_limits_shared[
+                    layer_idx] != 0:
+            swiglu_limit_shared = config.swiglu_limits_shared[layer_idx]
+        else:
+            swiglu_limit_shared = None
+        if config.swiglu_limits and config.swiglu_limits[
+                layer_idx] is not None and config.swiglu_limits[layer_idx] != 0:
+            swiglu_limit = config.swiglu_limits[layer_idx]
+        else:
+            swiglu_limit = None
+        if self.is_moe_layer:
+            self.moe = Step3p5MoEMLP(config, swiglu_limit=swiglu_limit)  #
+            self.share_expert = Step3p5MLP(
+                config,
+                intermediate_size=config.share_expert_dim,
+                swiglu_limit=swiglu_limit_shared)
+            self.use_moe = True
+        else:
+            self.mlp = Step3p5MLP(config,
+                                 intermediate_size=config.intermediate_size,
+                                 swiglu_limit=swiglu_limit_shared)
+
+        self.input_layernorm = Step3p5RMSNorm(
+            config.hidden_size,
+            eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Step3p5RMSNorm(
+            config.hidden_size,
+            eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[tuple[torch.Tensor]] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> torch.FloatTensor:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            cache_position=cache_position,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+        
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        if self.use_moe:
+            share_output = self.share_expert(hidden_states)
+            moe_output = self.moe(hidden_states)
+            ffn_output = moe_output + share_output
+        else:
+            ffn_output = self.mlp(hidden_states)
+        if isinstance(ffn_output, tuple):
+            hidden_states, _ = ffn_output
+        else:
+            hidden_states = ffn_output
+
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class Step3p5PreTrainedModel(PreTrainedModel):
+    # Link this model family to its configuration class so PreTrainedModel.from_pretrained
+    # can load the config instead of failing with a NoneType error.
+    config_class = Step3p5Config
+    supports_gradient_checkpointing = True
+    _skip_keys_device_placement = ["past_key_values"]
+    _keys_to_ignore_on_load_unexpected = [
+        r"model\.layers\.45\.*",
+        r"model\.layers\.46\.*",
+        r"model\.layers\.47\.*"
+    ]
+    _supports_flash_attn = False
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_static_cache = True
+    _supports_attention_backend = True
+
+
+class Step3p5Model(Step3p5PreTrainedModel, GenerationMixin):
+    _no_split_modules = ["Step3p5DecoderLayer"]
+    base_model_prefix = "model"
+    _tied_weights_keys = ["lm_head.weight"]
+    config: Step3p5Config
+    def __init__(self, config: Step3p5Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size,
+                                         self.padding_idx)
+        self.layers = nn.ModuleList([
+            Step3p5DecoderLayer(config, layer_idx)
+            for layer_idx in range(config.num_hidden_layers)
+        ])
+        self.norm = Step3p5RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.gradient_checkpointing = False
+        self.has_sliding_layers = "sliding_attention" in self.config.layer_types
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self, input_ids):
+        return self.embed_tokens(input_ids)
+
+    @can_return_tuple
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (output_hidden_states
+                                if output_hidden_states is not None else
+                                self.config.output_hidden_states)
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(
+                input_ids.to(self.embed_tokens.weight.device))
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length(
+            ) if past_key_values is not None else 0
+            cache_position = torch.arange(past_seen_tokens,
+                                          past_seen_tokens +
+                                          inputs_embeds.shape[1],
+                                          device=inputs_embeds.device)
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        hidden_states = inputs_embeds
+
+        # It may already have been prepared by e.g. `generate`
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            # Prepare mask arguments
+            mask_kwargs = {
+                "config": self.config,
+                "input_embeds": inputs_embeds,
+                "attention_mask": attention_mask,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "position_ids": position_ids,
+            }
+            # Create the masks
+            causal_mask_mapping = {
+                "full_attention": create_causal_mask(**mask_kwargs),
+            }
+
+            # The sliding window alternating layers are not always activated depending on the config
+            if self.has_sliding_layers:
+                causal_mask_mapping[
+                    "sliding_attention"] = create_sliding_window_causal_mask(
+                        **mask_kwargs)
+
+        # # create position embeddings to be shared across the decoder layers
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        for decoder_layer in self.layers[:self.config.num_hidden_layers]:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states, )
+
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=causal_mask_mapping[
+                    decoder_layer.attention_type],
+                position_ids=position_ids,
+                past_key_value=past_key_values,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+            hidden_states = layer_outputs
+
+        hidden_states = self.norm(hidden_states)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class Step3p5ForCausalLM(Step3p5PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    config: Step3p5Config
+
+    def __init__(self, config: Step3p5Config):
+        super().__init__(config)
+        self.model = Step3p5Model(config)
+        self.lm_head = nn.Linear(config.hidden_size,
+                                 config.vocab_size,
+                                 bias=False)
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.model.set_input_embeddings(value)
+
+    def get_output_embeddings(self):
+        return self.model.get_output_embeddings()
+
+    def set_output_embeddings(self, new_embeddings):
+        self.model.set_output_embeddings(new_embeddings)
+
+    def set_decoder(self, decoder):
+        self.model.set_decoder(decoder)
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+    
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        num_patches=None,
+        patch_pixel_values=None,
+        patch_newline_mask=None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> Union[tuple, Step3p5CausalLMOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+        Example:
+        ```python
+        >>> from transformers import AutoTokenizer, Llama4ForCausalLM
+        >>> model = Llama4ForCausalLM.from_pretrained("meta-llama4/Llama4-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama4/Llama4-2-7b-hf")
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (output_hidden_states
+                                if output_hidden_states is not None else
+                                self.config.output_hidden_states)
+        # breakpoint()
+        outputs = self.model(
+            input_ids=input_ids,
+            num_patches=num_patches,
+            patch_pixel_values=patch_pixel_values,
+            patch_newline_mask=patch_newline_mask,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            **kwargs,
+        )
+        hidden_states = outputs.last_hidden_state
+        logits = self.lm_head(hidden_states)
+
+        return Step3p5CausalLMOutputWithPast(logits=logits, )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        inputs_embeds=None,
+        pixel_values=None,
+        attention_mask=None,
+        cache_position=None,
+        logits_to_keep=None,
+        **kwargs,
+    ):
+
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+            **kwargs,
+        )
+
+        if cache_position[0] == 0:
+            # If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
+            # Otherwise we need pixel values to be passed to model
+            model_inputs["pixel_values"] = pixel_values
+
+        return model_inputs
+
+    def _fix_state_dict_key_on_load(self, key: str) -> tuple[str, bool]:
+        if key.startswith("language_model."):
+            return key[len("language_model."):], True
+
+        return key, False

special_tokens_map.json

@@ -0,0 +1,17 @@
+{
+  "bos_token": {
+    "content": "<｜begin▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|im_end|>"
+}

step3p5_quantize_quark.py

@@ -0,0 +1,566 @@
+#!/usr/bin/env python3
+#
+# Copyright (C) 2023 - 2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+#
+# Quantization script for Step-3.5-Flash with MoE layer replacement
+
+from __future__ import annotations
+
+import argparse
+import os
+import re
+import shutil
+from pathlib import Path
+from types import MethodType
+from typing import Any
+
+import torch
+import torch.nn as nn
+
+from quark.torch import LLMTemplate, ModelQuantizer, export_safetensors
+from quark.torch.utils.llm import (
+    get_calib_dataloader,
+    get_model,
+    get_tokenizer,
+)
+from quark.common.utils.log import ScreenLogger
+
+try:
+    # Needed only when the model is loaded with accelerate offload (meta tensors).
+    from accelerate.hooks import AlignDevicesHook, add_hook_to_module  # type: ignore
+    from accelerate.utils import PrefixedDataset  # type: ignore
+
+    _ACCELERATE_AVAILABLE = True
+except Exception:
+    AlignDevicesHook = None  # type: ignore[assignment]
+    add_hook_to_module = None  # type: ignore[assignment]
+    PrefixedDataset = None  # type: ignore[assignment]
+    _ACCELERATE_AVAILABLE = False
+
+
+DEFAULT_INPUT_MODEL_PATH = "stepfun-ai/Step-3.5-Flash"
+DEFAULT_OUTPUT_MODEL_PATH = "quantized_models/Step-3.5-Flash-MXFP4"
+
+logger = ScreenLogger(__name__)
+
+def _step35_template_exclude_layers() -> list[str]:
+    return [
+        # embeddings / lm head / norms
+        "model.embed_tokens*",
+        "*embed_tokens*",
+        "*lm_head*",
+        "*layernorm*",
+        "*norm*",
+        # Router gate
+        "*moe.gate",
+        "*moe.router_bias*",
+        # The first three blocks use dense FFNs
+        "model.layers.0.mlp.*",
+        "model.layers.1.mlp.*",
+        "model.layers.2.mlp.*",
+        # Shared Experts
+        "*share_expert*",
+        "*self_attn*",
+    ]
+
+PRESETS: dict[str, dict[str, Any]] = {
+
+    "mxfp4_moe_only_no_kvcache": {
+        "quant_scheme": "mxfp4",
+        "exclude_layers": _step35_template_exclude_layers(),
+    },
+}
+
+
+def _copy_non_weight_files(src_dir: str, dst_dir: str) -> None:
+    """
+    Copy non-weight files from an HF model directory (json/jinja/tokenizer, etc.),
+    while skipping *.safetensors and model.safetensors.index.json.
+    
+    Note: `export_safetensors` exports the essential HF weights and config, but the
+    original model directory may contain extra assets (e.g. chat_template.jinja).
+    We do a conservative copy here so offline inference keeps those auxiliary files.
+    """
+    src = Path(src_dir)
+    dst = Path(dst_dir)
+    dst.mkdir(parents=True, exist_ok=True)
+
+    for p in src.iterdir():
+        if p.is_dir():
+            continue
+        name = p.name
+        if name.endswith(".safetensors"):
+            continue
+        if name == "model.safetensors.index.json":
+            continue
+        # Export will (re-)write config / generation_config; copying them here is harmless
+        # (later writes will overwrite).
+        shutil.copy2(p, dst / name)
+
+
+def _register_step35_flash_template() -> None:
+    """
+    Register a Quark LLMTemplate for Step-3.5-Flash (config.model_type = step3p5).
+    """
+    model_type = "step3p5"
+    if model_type in LLMTemplate.list_available():
+        return
+
+
+    step35_flash_template = LLMTemplate(
+        model_type=model_type,
+        kv_layers_name=["*k_proj", "*v_proj"],
+        q_layer_name="*q_proj",
+        exclude_layers_name=_step35_template_exclude_layers(),
+    )
+    LLMTemplate.register_template(step35_flash_template)
+    logger.info("Registered LLMTemplate: %s", model_type)
+
+
+@torch.no_grad()
+def replace_step35_moelinear_with_linear(moe_module: Any) -> None:
+    """
+    Convert Step3p5MoEMLP's MoELinear modules into separate Linear layers per expert.
+    """
+    if getattr(moe_module, "_step35_replaced", False):
+        return
+
+    logger.debug("Converting Step3p5MoEMLP experts to separate gate/up/down Linear layers...")
+
+    # Get dimensions from the module
+    num_experts: int = int(getattr(moe_module, "moe_num_experts", 288))
+    hidden_size: int = int(getattr(moe_module, "hidden_size", 4096))
+    moe_intermediate_size: int = int(getattr(moe_module, "moe_intermediate_size", 1280))
+    
+    # Store original device and dtype from one of the MoELinear modules
+    original_device = moe_module.gate_proj.weight.device
+    original_dtype = moe_module.gate_proj.weight.dtype
+    # [num_experts, in, out]
+    # Expose common attribute names for the forward helper
+    moe_module.hidden_size = hidden_size
+    moe_module.expert_dim = moe_intermediate_size
+    moe_module.num_experts = num_experts
+
+    is_meta: bool = original_device == torch.device("meta")
+    target_device_for_new = original_device if not is_meta else torch.device("meta")
+
+    # Create individual expert modules, each containing gate_proj, up_proj, down_proj
+    for expert_index in range(num_experts):
+        expert_module = nn.Module()
+        expert_module.gate_proj = nn.Linear(
+            hidden_size, moe_intermediate_size, bias=False, device=target_device_for_new, dtype=original_dtype
+        )
+        expert_module.up_proj = nn.Linear(
+            hidden_size, moe_intermediate_size, bias=False, device=target_device_for_new, dtype=original_dtype
+        )
+        expert_module.down_proj = nn.Linear(
+            moe_intermediate_size, hidden_size, bias=False, device=target_device_for_new, dtype=original_dtype
+        )
+        setattr(moe_module, str(expert_index), expert_module)
+
+
+    # Sync weights from MoELinear to individual Linear modules
+    weights_synced = _step35_sync_weights_to_linear(moe_module)
+    
+    # Replace forward method
+    moe_module.forward = MethodType(_step35_moe_forward, moe_module)
+
+    if weights_synced:
+        _step35_cleanup_fused(moe_module)
+
+    moe_module._step35_replaced = True
+
+
+@torch.no_grad()
+def _step35_sync_weights_to_linear(module: Any) -> bool:
+    """
+    Split MoELinear weights and copy into per-expert Linear layers.
+    Returns True if synced; returns False if fused weights are still on 'meta' (not materialized).
+    MoELinear tensors in Step3p5MoEMLP are expected to be:
+      - gate_proj.weight: [num_experts, moe_intermediate_size, hidden_size]
+      - up_proj.weight:   [num_experts, moe_intermediate_size, hidden_size]
+      - down_proj.weight: [num_experts, hidden_size, moe_intermediate_size]
+    """
+    if getattr(module, "_weights_synced", False):
+        return True
+
+    W_gate = getattr(module, "gate_proj", None)
+    W_up = getattr(module, "up_proj", None)
+    W_down = getattr(module, "down_proj", None)
+    
+    if W_gate is None or W_up is None or W_down is None:
+        return False
+
+    is_offload = getattr(W_gate.weight, "is_meta", False) or W_gate.weight.device == torch.device("meta")
+    if is_offload:
+        # Loaded with accelerate offload: tensors live in module._hf_hook.weights_map on CPU.
+        if not _ACCELERATE_AVAILABLE:
+            raise RuntimeError(
+                "Model appears to be loaded with accelerate offload (meta tensors), but accelerate is not available."
+            )
+        if not hasattr(module, "_hf_hook"):
+            return False
+        W_gate = module._hf_hook.weights_map["gate_proj.weight"]
+        W_up = module._hf_hook.weights_map["up_proj.weight"]
+        W_down = module._hf_hook.weights_map["down_proj.weight"]
+
+    try:
+        for expert_index in range(int(module.num_experts)):
+            expert_module = getattr(module, str(expert_index))
+
+            W_gate_current = W_gate.weight[expert_index]  # [moe_intermediate_size, hidden_size]
+            W_up_current = W_up.weight[expert_index]      # [moe_intermediate_size, hidden_size]
+            W_down_current = W_down.weight[expert_index]  # [hidden_size, moe_intermediate_size]
+
+            if is_offload:
+                hook = module._hf_hook
+                dataset = hook.weights_map.dataset
+                layer_value = [W_gate_current, W_up_current, W_down_current]
+                for idx, layer_name in enumerate(["gate_proj", "up_proj", "down_proj"]):
+                    prefix = f"{hook.weights_map.prefix}{expert_index}.{layer_name}."
+                    prefixed_weights_map = PrefixedDataset(dataset, prefix)
+                    full_name = f"{prefix}weight"
+                    dataset.all_keys.append(full_name)
+                    dataset.state_dict[full_name] = layer_value[idx]
+
+                    quark_hook = AlignDevicesHook(
+                        execution_device=hook.execution_device,
+                        offload=hook.offload,
+                        io_same_device=hook.io_same_device,
+                        weights_map=prefixed_weights_map,
+                        offload_buffers=hook.offload_buffers,
+                        place_submodules=hook.place_submodules,
+                        skip_keys=hook.skip_keys,
+                        tied_params_map=hook.tied_params_map,
+                    )
+                    linear_module = getattr(expert_module, layer_name)
+                    add_hook_to_module(linear_module, quark_hook)
+            else:
+                # No transpose needed: nn.Linear expects [out_features, in_features], which matches MoELinear tensors.
+                expert_module.gate_proj.weight.data.copy_(W_gate_current.to(W_gate.weight.device))
+                expert_module.up_proj.weight.data.copy_(W_up_current.to(W_up.weight.device))
+                expert_module.down_proj.weight.data.copy_(W_down_current.to(W_down.weight.device))
+
+        if is_offload:
+            prefix = module._hf_hook.weights_map.prefix
+            del module._hf_hook.weights_map.dataset.state_dict[f"{prefix}gate_proj.weight"]
+            del module._hf_hook.weights_map.dataset.state_dict[f"{prefix}up_proj.weight"]
+            del module._hf_hook.weights_map.dataset.state_dict[f"{prefix}down_proj.weight"]
+            module._hf_hook.weights_map.dataset.all_keys.remove(f"{prefix}gate_proj.weight")
+            module._hf_hook.weights_map.dataset.all_keys.remove(f"{prefix}up_proj.weight")
+            module._hf_hook.weights_map.dataset.all_keys.remove(f"{prefix}down_proj.weight")
+
+        module._weights_synced = True
+        return True
+    except Exception as e:
+        logger.warning("Failed to sync Step3.5 MoE weights: %s", e)
+        return False
+
+
+
+@torch.no_grad()
+def _step35_cleanup_fused(module: Any) -> None:
+    """Optionally remove fused MoELinear modules after replacement."""
+    # The original MoELinear modules should be garbage collected
+    # when they're replaced, but we can explicitly clear references
+    for proj_name in ["gate_proj", "up_proj", "down_proj"]:
+        # Clear any remaining references to original MoELinear
+        if hasattr(module, proj_name):
+            delattr(module, proj_name)
+    
+    torch.cuda.empty_cache()
+    logger.debug(f"Cleaned up original MoELinear modules")
+
+
+def _step35_moe_forward(self: Any, hidden_states: torch.Tensor) -> torch.Tensor:
+    """
+    Forward using per-expert gate_proj, up_proj, down_proj (nn.Linear),
+    matching the original Step3p5MoEMLP.forward semantics but without MoELinear.
+    """
+    synced = _step35_sync_weights_to_linear(self)
+    if not synced:
+        raise RuntimeError(
+            "Step3p5MoEMLP weights are on 'meta' (not materialized). "
+            "Move fused parameters to a real device first, then call forward."
+        )
+    
+    batch_size, sequence_length, hidden_dim = hidden_states.shape
+    hidden_states = hidden_states.view(-1, hidden_dim)
+    
+    # Router/gating
+    if self.need_fp32_gate:
+        router_logits = torch.matmul(hidden_states.to(torch.float32), self.gate.weight.t().to(torch.float32))
+    else:
+        # router_logits: (batch * sequence_length, n_experts)
+        router_logits = self.gate(hidden_states)
+    
+    # Custom routing or standard softmax + top-k
+    if hasattr(self, 'custom_routing_function') and self.custom_routing_function:
+        routing_weights, selected_experts = self.custom_routing_function(
+            router_logits, self.top_k, renormalize=True)
+    else:
+        routing_weights = torch.nn.functional.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+    
+    # Apply scaling factor
+    routing_weights = routing_weights * self.routed_scaling_factor
+    
+    # Initialize output
+    final_hidden_states = torch.zeros(
+        (batch_size * sequence_length, hidden_dim),
+        dtype=hidden_states.dtype,
+        device=hidden_states.device)
+    
+    # One hot encode the selected experts to create an expert mask
+    # this will be used to easily index which expert is going to be solicited
+    expert_mask = torch.nn.functional.one_hot(
+        selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
+    
+    limit = getattr(self, 'limit', None)
+    
+    # Loop over all available experts in the model and perform the computation on each expert
+    for expert_idx in range(self.num_experts):
+        idx, top_x = torch.where(expert_mask[expert_idx])
+
+        # Index the correct hidden states and compute the expert hidden state for
+        # the current expert. We need to make sure to multiply the output hidden
+        # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+        current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+        
+        expert_module = getattr(self, str(expert_idx))
+
+        up = expert_module.up_proj(current_state)
+        gate = self.act_fn(expert_module.gate_proj(current_state))
+        
+        if limit is not None:
+            gate = gate.clamp(min=None, max=limit)
+            up = up.clamp(min=-limit, max=limit)
+        
+        current_hidden_states = expert_module.down_proj(gate * up) * routing_weights[top_x, idx, None]
+        
+        # However `index_add_` only support torch tensors for indexing so we'll use
+        # the `top_x` tensor here.
+        final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+    
+    final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+    return final_hidden_states
+
+@torch.no_grad()
+def patch_step35_moe(model: nn.Module) -> int:
+    """
+    Apply Step-3.5-Flash MoE replacement to all Step3p5MoEMLP modules in the model.
+    """
+    patched = 0
+    for name, module in model.named_modules(remove_duplicate=False):
+        if module.__class__.__name__ == "Step3p5MoEMLP":
+            replace_step35_moelinear_with_linear(module)
+            patched += 1
+            logger.debug(f"Patched MoE module: {name}")
+
+    if patched > 0:
+        logger.info("Patched %d Step3p5MoEMLP module(s) for quantization.", patched)
+    return patched
+
+
+def _resolve_calib_device(device: str, model: nn.Module) -> str:
+    """
+    Resolve a torch-compatible device string for calibration inputs.
+    """
+    if device != "auto":
+        return str(device)
+
+    hf_map = getattr(model, "hf_device_map", None)
+    if isinstance(hf_map, dict):
+        cuda_ids: list[int] = []
+        for v in hf_map.values():
+            m = re.match(r"^cuda:(\d+)$", str(v))
+            if m:
+                cuda_ids.append(int(m.group(1)))
+        if cuda_ids:
+            return f"cuda:{min(cuda_ids)}"
+
+    if torch.cuda.is_available():
+        return "cuda:0"
+    return "cpu"
+
+
+def main(args: argparse.Namespace) -> None:
+    os.makedirs(args.output_quantized_hf_path, exist_ok=True)
+
+    _register_step35_flash_template()
+
+    if getattr(args, "preset", None):
+        preset_cfg = PRESETS[args.preset]
+        args.quant_scheme = preset_cfg["quant_scheme"]
+        if getattr(args, "quant_algo", None) is None and "quant_algo" in preset_cfg:
+            args.quant_algo = preset_cfg["quant_algo"]
+        logger.info("Using preset: %s", args.preset)
+
+    logger.info("Input model: %s", args.model_dir)
+    logger.info("Output dir: %s", args.output_quantized_hf_path)
+
+    logger.info("Step 1/4: Loading model and tokenizer ...")
+    model, _ = get_model(
+        args.model_dir,
+        data_type=args.data_type,
+        device=args.device,
+        multi_gpu=args.multi_gpu,
+        multi_device=args.multi_device,
+        attn_implementation=args.model_attn_implementation,
+        trust_remote_code=args.trust_remote_code,
+    )
+    
+    patch_step35_moe(model)
+    
+    model_type = model.config.model_type if hasattr(model.config, "model_type") else model.config.architectures[0]
+    tokenizer = get_tokenizer(
+        args.model_dir, max_seq_len=args.seq_len, model_type=model_type, trust_remote_code=args.trust_remote_code
+    )
+
+    logger.info("Step 2/4: Building calibration dataloader ...")
+    base_device = str(model.device) if (args.multi_gpu or args.multi_device) else str(args.device)
+    main_device = _resolve_calib_device(base_device, model)
+    logger.info("Calibration dataset: %s", args.dataset)
+    calib_dataloader = get_calib_dataloader(
+        dataset_name=args.dataset,
+        tokenizer=tokenizer,
+        batch_size=args.batch_size,
+        num_calib_data=args.num_calib_data,
+        seqlen=args.seq_len,
+        device=main_device,
+    )
+
+    logger.info("Step 3/4: Quantizing ...")
+    template = LLMTemplate.get(model_type)
+    if args.exclude_layers is not None:
+        logger.warning(
+            "Ignoring --exclude_layers (%s). This script always uses "
+            "_register_step35_flash_template excludes for Step-3.5-Flash.",
+            args.exclude_layers,
+        )
+    exclude_layers = _step35_template_exclude_layers()
+    logger.info("Exclude layers (template): %s", exclude_layers)
+    if getattr(args, "quant_algo", None):
+        logger.info("Quantization algorithm(s): %s", args.quant_algo)
+    
+    quant_config = template.get_config(
+        scheme=args.quant_scheme,
+        algorithm=args.quant_algo,
+        kv_cache_scheme=None,
+        min_kv_scale=0.0,
+        layer_config={},
+        attention_scheme=None,
+        exclude_layers=exclude_layers,
+        algo_configs=None,
+    )
+    
+    quantizer = ModelQuantizer(quant_config, args.multi_device)
+    model = quantizer.quantize_model(model, calib_dataloader)
+    
+    model = quantizer.freeze(model)
+
+    logger.info("Step 4/4: Exporting HF safetensors ...")
+    _copy_non_weight_files(args.model_dir, args.output_quantized_hf_path)
+    with torch.no_grad():
+        export_safetensors(
+            model=model,
+            output_dir=args.output_quantized_hf_path,
+            custom_mode="quark",
+            weight_format=args.export_weight_format,
+            pack_method=args.pack_method,
+        )
+        tokenizer.save_pretrained(args.output_quantized_hf_path)
+
+    logger.info("Export completed.")
+    logger.info("========== Quantization Completed Successfully ==========")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Offline quantization for Step-3.5-Flash with MoE layer replacement"
+    )
+    parser.add_argument("--model_dir", dest="model_dir", type=str, default=DEFAULT_INPUT_MODEL_PATH)
+    parser.add_argument("--output_dir", dest="output_quantized_hf_path", type=str, default=DEFAULT_OUTPUT_MODEL_PATH)
+
+    # Model loading
+    parser.add_argument("--device", type=str, default="auto", choices=["auto", "cuda", "cpu"])
+    parser.add_argument("--multi_gpu", dest="multi_gpu", action="store_true")
+    parser.add_argument("--multi_device", dest="multi_device", action="store_true")
+    parser.add_argument(
+        "--model_attn_implementation",
+        dest="model_attn_implementation",
+        type=str,
+        default="eager",
+        choices=["eager", "sdpa", "flash_attention_2"],
+    )
+    parser.add_argument(
+        "--data_type",
+        dest="data_type",
+        type=str,
+        default="auto",
+        choices=["auto", "float16", "bfloat16", "float32"],
+    )
+
+    # Calibration
+    parser.add_argument(
+        "--dataset",
+        dest="dataset",
+        type=str,
+        default="pileval",
+        help="Calibration dataset name. Default is 'pileval'.",
+    )
+    parser.add_argument("--seq_len", dest="seq_len", type=int, default=512)
+    parser.add_argument("--batch_size", dest="batch_size", type=int, default=1)
+    parser.add_argument("--num_calib_data", dest="num_calib_data", type=int, default=128)
+
+    # Quantization
+    parser.add_argument(
+        "--preset",
+        dest="preset",
+        type=str,
+        choices=sorted(PRESETS.keys()),
+        default="mxfp4_moe_only_no_kvcache",
+        help="Convenience preset for quantization settings.",
+    )
+    parser.add_argument(
+        "--quant_algo",
+        dest="quant_algo",
+        type=str,
+        default=None,
+        help="Optional quantization algorithm(s) to apply.",
+    )
+    parser.add_argument(
+        "--exclude_layers",
+        type=str,
+        nargs="*",
+        default=None,
+        help="Layer wildcard patterns to exclude from quantization.",
+    )
+
+    # Export
+    parser.add_argument("--pack_method", dest="pack_method", type=str, default="reorder", choices=["order", "reorder"])
+    parser.add_argument(
+        "--export_weight_format",
+        dest="export_weight_format",
+        type=str,
+        default="real_quantized",
+        choices=["fake_quantized", "real_quantized"],
+    )
+    group = parser.add_mutually_exclusive_group()
+    group.add_argument(
+        "--trust_remote_code",
+        action="store_true",
+        dest="trust_remote_code",
+        help="Enable execution of custom model code from the Hub (use only with repositories you fully trust).",
+    )
+    group.add_argument(
+        "--no_trust_remote_code",
+        action="store_false",
+        dest="trust_remote_code",
+        help="Disable execution of custom model code from the Hub (safer, recommended if unsure).",
+    )
+    parser.set_defaults(trust_remote_code=True)
+
+    main(parser.parse_args())
+