diff --git a/.gitattributes b/.gitattributes index a6344aac8c09253b3b630fb776ae94478aa0275b..a09db2ea4d1bd1fc1c09f6fca45e8e4054953535 100644 --- a/.gitattributes +++ b/.gitattributes @@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text *.zip filter=lfs diff=lfs merge=lfs -text *.zst filter=lfs diff=lfs merge=lfs -text *tfevents* filter=lfs diff=lfs merge=lfs -text +tokenizer.json filter=lfs diff=lfs merge=lfs -text +model.safetensors.index.json filter=lfs diff=lfs merge=lfs -text diff --git a/__init__.py b/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..fef8d30e777e13534cecaa58eb361f6de857247b --- /dev/null +++ b/__init__.py @@ -0,0 +1,4 @@ +from .configuration_afmoe import AfmoeConfig +from .modeling_afmoe import AfmoeForCausalLM + +__all__ = ["AfmoeConfig", "AfmoeForCausalLM"] diff --git a/chat_template.jinja b/chat_template.jinja new file mode 100644 index 0000000000000000000000000000000000000000..5b93512b2ca66f53ececeff7a4661df3a4084118 --- /dev/null +++ b/chat_template.jinja @@ -0,0 +1,159 @@ +<|begin_of_text|>{%- macro render_extra_keys(json_dict, handled_keys) -%} + {%- if json_dict is mapping %} + {%- for json_key in json_dict if json_key not in handled_keys %} + {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %} + {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '' }} + {%- else %} + {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '' }} + {%- endif %} + {%- endfor %} + {%- endif %} +{%- endmacro -%} + +{%- macro render_tool_call(raw_tool_call) -%} + {%- if raw_tool_call.function is defined and raw_tool_call.function is mapping %} + {%- set tool_call = raw_tool_call.function %} + {%- else %} + {%- set tool_call = raw_tool_call %} + {%- endif %} + {{- '\n\n' }} + {%- if tool_call.arguments is defined and tool_call.arguments is mapping %} + {%- for args_name, args_value in tool_call.arguments.items() %} + {{- '\n' }} + {%- if args_value is mapping or (args_value is sequence and args_value is not string) %} + {{- args_value | tojson | safe }} + {%- else %} + {{- args_value | string }} + {%- endif %} + {{- '\n\n' }} + {%- endfor %} + {%- endif %} + {{- '\n' }} +{%- endmacro -%} + +{%- set system_message = none %} +{%- if messages and messages[0]["role"] == "system" %} + {%- set system_message = messages[0]["content"] %} + {%- set loop_messages = messages[1:] %} +{%- else %} + {%- set loop_messages = messages %} +{%- endif %} + +{%- if not tools is defined %} + {%- set tools = [] %} +{%- endif %} +{%- set has_tools = tools is iterable and tools is not string and tools | length > 0 %} + +{%- if system_message is not none or has_tools %} + {{- '<|im_start|>system\n' }} + {%- if system_message is not none %} + {{- system_message }} + {%- else %} + {{- "You are Trinity Large, a helpful assistant developed by Arcee AI, that can interact with a computer to solve tasks." }} + {%- endif %} + {%- if has_tools %} + {{- "\n\n# Tools\n\nYou have access to the following functions:\n\n" }} + {%- for tool in tools %} + {%- if tool.function is defined and tool.function is mapping %} + {%- set tool = tool.function %} + {%- endif %} + {{- '\n\n' ~ (tool.name | default('') | string) ~ '' }} + {%- if tool.description is defined and tool.description is not none %} + {{- '\n' ~ (tool.description | string | trim) ~ '' }} + {%- endif %} + {{- '\n' }} + {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %} + {%- for param_name, param_fields in tool.parameters.properties.items() %} + {{- '\n\n' ~ (param_name | string) ~ '' }} + {%- if param_fields is mapping and param_fields.type is defined and param_fields.type is not none %} + {{- '\n' ~ (param_fields.type | string) ~ '' }} + {%- endif %} + {%- if param_fields is mapping and param_fields.description is defined and param_fields.description is not none %} + {{- '\n' ~ (param_fields.description | string | trim) ~ '' }} + {%- endif %} + {%- if param_fields is mapping %} + {%- set handled_keys = ['name', 'type', 'description'] %} + {{- render_extra_keys(param_fields, handled_keys) }} + {%- endif %} + {{- '\n' }} + {%- endfor %} + {%- endif %} + {%- if tool.parameters is defined %} + {%- set handled_keys = ['type', 'properties'] %} + {{- render_extra_keys(tool.parameters, handled_keys) }} + {%- endif %} + {{- '\n' }} + {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %} + {{- render_extra_keys(tool, handled_keys) }} + {{- '\n' }} + {%- endfor %} + {{- "\n" }} + {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n\n\n\nvalue_1\n\n\nThis is the value for the second parameter\nthat can span\nmultiple lines\n\n\n\n\n\nReminder:\n- Function calls MUST follow the specified format: an inner block must be nested within XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n' }} + {%- endif %} + {{- '<|im_end|>\n' }} +{%- endif %} + +{%- for message in loop_messages %} + {%- set role = message.role | default('') %} + {%- if role == "assistant" %} + {%- set content_str = '' if message.content is none else (message.content | string) %} + {%- set trimmed_content = content_str | trim %} + + {%- set has_reasoning_content = message.reasoning_content is defined %} + {%- set has_reasoning = has_reasoning_content or (message.reasoning is defined) %} + + {%- if has_reasoning_content %} + {%- set reasoning_value = message.reasoning_content %} + {%- elif message.reasoning is defined %} + {%- set reasoning_value = message.reasoning %} + {%- else %} + {%- set reasoning_value = none %} + {%- endif %} + + {%- set has_tool_calls = message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls is not string and message.tool_calls | length > 0 %} + + {{- '<|im_start|>assistant\n' }} + {%- if has_reasoning %} + {%- if reasoning_value %} + {{- '' + (reasoning_value | string | trim) + '' }} + {%- else %} + {{- '' }} + {%- endif %} + {%- if trimmed_content %} + {{- '\n' + trimmed_content }} + {%- endif %} + {%- elif has_tool_calls %} + {%- if trimmed_content %} + {{- trimmed_content }} + {%- endif %} + {%- else %} + {{- content_str }} + {%- endif %} + + {%- if has_tool_calls %} + {%- for tool_call in message.tool_calls %} + {%- set separator = '\n' if ((loop.first and (has_reasoning or trimmed_content)) or (not loop.first)) else '' -%} + {{- separator + render_tool_call(tool_call) }} + {%- endfor %} + {%- endif %} + {{- '<|im_end|>\n' }} + {%- elif role == "tool" or role == "observation" or role == "function" %} + {%- if loop.first or loop.previtem.role not in ["tool", "observation", "function"] %} + {{- '<|im_start|>user\n' }} + {%- endif %} + {{- '\n' }} + {{- '' if message.content is none else (message.content | string) }} + {{- '\n\n' }} + {%- if loop.last or loop.nextitem.role not in ["tool", "observation", "function"] %} + {{- '<|im_end|>\n' }} + {%- endif %} + {%- else %} + {{- '<|im_start|>' + (role | string) }} + {{- '\n' + ('' if message.content is none else (message.content | string)) }} + {{- '<|im_end|>\n' }} + {%- endif %} +{%- endfor %} + +{%- if add_generation_prompt %} + {{- '<|im_start|>assistant\n' }} +{%- endif %} \ No newline at end of file diff --git a/config.json b/config.json new file mode 100644 index 0000000000000000000000000000000000000000..7c3ed65eebfa7c377a2a53e4147881c0f4b9fe5a --- /dev/null +++ b/config.json @@ -0,0 +1,512 @@ +{ + "architectures": [ + "AfmoeForCausalLM" + ], + "attention_dropout": 0.0, + "auto_map": { + "AutoConfig": "configuration_afmoe.AfmoeConfig", + "AutoModel": "modeling_afmoe.AfmoeModel", + "AutoModelForCausalLM": "modeling_afmoe.AfmoeForCausalLM" + }, + "dtype": "bfloat16", + "global_attn_every_n_layers": 4, + "head_dim": 128, + "hidden_act": "silu", + "hidden_size": 3072, + "initializer_range": 0.02, + "intermediate_size": 12288, + "layer_types": [ + "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", + "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" + ], + "load_balance_coeff": 5e-05, + "max_position_embeddings": 262144, + "model_type": "afmoe", + "moe_intermediate_size": 3072, + "mup_enabled": true, + "n_group": 1, + "num_attention_heads": 48, + "num_dense_layers": 6, + "num_expert_groups": 1, + "num_experts": 256, + "num_experts_per_tok": 4, + "num_hidden_layers": 60, + "num_key_value_heads": 8, + "num_limited_groups": 1, + "num_shared_experts": 1, + "quantization_config": { + "config_groups": { + "group_0": { + "format": "nvfp4-pack-quantized", + "input_activations": { + "actorder": null, + "block_structure": null, + "dynamic": "local", + "group_size": 16, + "num_bits": 4, + "observer": "static_minmax", + "observer_kwargs": {}, + "scale_dtype": "torch.float8_e4m3fn", + "strategy": "tensor_group", + "symmetric": true, + "type": "float", + "zp_dtype": null + }, + "output_activations": null, + "targets": [ + "Linear" + ], + "weights": { + "actorder": null, + "block_structure": null, + "dynamic": false, + "group_size": 16, + "num_bits": 4, + "observer": "memoryless_minmax", + "observer_kwargs": {}, + "scale_dtype": "torch.float8_e4m3fn", + "strategy": "tensor_group", + "symmetric": true, + "type": "float", + "zp_dtype": null + } + } + }, + "format": "nvfp4-pack-quantized", + "global_compression_ratio": null, + "ignore": [ + "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.gate_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.gate_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.gate_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.gate_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.gate_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.gate_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.gate_proj", + "model.layers.6.mlp.router.gate", + "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.gate_proj", + "model.layers.7.mlp.router.gate", + "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.gate_proj", + "model.layers.8.mlp.router.gate", + "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.gate_proj", + "model.layers.9.mlp.router.gate", + "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.gate_proj", + "model.layers.10.mlp.router.gate", + "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.gate_proj", + "model.layers.11.mlp.router.gate", + "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.gate_proj", + "model.layers.12.mlp.router.gate", + "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.gate_proj", + "model.layers.13.mlp.router.gate", + "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.gate_proj", + "model.layers.14.mlp.router.gate", + "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.gate_proj", + "model.layers.15.mlp.router.gate", + "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.gate_proj", + "model.layers.16.mlp.router.gate", + "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.gate_proj", + "model.layers.17.mlp.router.gate", + "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.gate_proj", + "model.layers.18.mlp.router.gate", + "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.gate_proj", + "model.layers.19.mlp.router.gate", + "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.gate_proj", + "model.layers.20.mlp.router.gate", + "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.gate_proj", + "model.layers.21.mlp.router.gate", + "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.gate_proj", + "model.layers.22.mlp.router.gate", + "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.gate_proj", + "model.layers.23.mlp.router.gate", + "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.gate_proj", + "model.layers.24.mlp.router.gate", + "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.gate_proj", + "model.layers.25.mlp.router.gate", + "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.gate_proj", + "model.layers.26.mlp.router.gate", + "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.gate_proj", + "model.layers.27.mlp.router.gate", + "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.gate_proj", + "model.layers.28.mlp.router.gate", + "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.gate_proj", + "model.layers.29.mlp.router.gate", + "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.gate_proj", + "model.layers.30.mlp.router.gate", + "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.gate_proj", + "model.layers.31.mlp.router.gate", + "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.gate_proj", + "model.layers.32.mlp.router.gate", + "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.gate_proj", + "model.layers.33.mlp.router.gate", + "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.gate_proj", + "model.layers.34.mlp.router.gate", + "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.gate_proj", + "model.layers.35.mlp.router.gate", + "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.gate_proj", + "model.layers.36.mlp.router.gate", + "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.gate_proj", + "model.layers.37.mlp.router.gate", + "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.gate_proj", + "model.layers.38.mlp.router.gate", + "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.gate_proj", + "model.layers.39.mlp.router.gate", + "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.gate_proj", + "model.layers.40.mlp.router.gate", + "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.gate_proj", + "model.layers.41.mlp.router.gate", + "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.gate_proj", + "model.layers.42.mlp.router.gate", + "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.gate_proj", + "model.layers.43.mlp.router.gate", + "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.gate_proj", + "model.layers.44.mlp.router.gate", + "model.layers.45.self_attn.q_proj", + "model.layers.45.self_attn.k_proj", + "model.layers.45.self_attn.v_proj", + "model.layers.45.self_attn.o_proj", + "model.layers.45.self_attn.gate_proj", + "model.layers.45.mlp.router.gate", + "model.layers.46.self_attn.q_proj", + "model.layers.46.self_attn.k_proj", + "model.layers.46.self_attn.v_proj", + "model.layers.46.self_attn.o_proj", + "model.layers.46.self_attn.gate_proj", + "model.layers.46.mlp.router.gate", + "model.layers.47.self_attn.q_proj", + "model.layers.47.self_attn.k_proj", + "model.layers.47.self_attn.v_proj", + "model.layers.47.self_attn.o_proj", + "model.layers.47.self_attn.gate_proj", + "model.layers.47.mlp.router.gate", + "model.layers.48.self_attn.q_proj", + "model.layers.48.self_attn.k_proj", + "model.layers.48.self_attn.v_proj", + "model.layers.48.self_attn.o_proj", + "model.layers.48.self_attn.gate_proj", + "model.layers.48.mlp.router.gate", + "model.layers.49.self_attn.q_proj", + "model.layers.49.self_attn.k_proj", + "model.layers.49.self_attn.v_proj", + "model.layers.49.self_attn.o_proj", + "model.layers.49.self_attn.gate_proj", + "model.layers.49.mlp.router.gate", + "model.layers.50.self_attn.q_proj", + "model.layers.50.self_attn.k_proj", + "model.layers.50.self_attn.v_proj", + "model.layers.50.self_attn.o_proj", + "model.layers.50.self_attn.gate_proj", + "model.layers.50.mlp.router.gate", + "model.layers.51.self_attn.q_proj", + "model.layers.51.self_attn.k_proj", + "model.layers.51.self_attn.v_proj", + "model.layers.51.self_attn.o_proj", + "model.layers.51.self_attn.gate_proj", + "model.layers.51.mlp.router.gate", + "model.layers.52.self_attn.q_proj", + "model.layers.52.self_attn.k_proj", + "model.layers.52.self_attn.v_proj", + "model.layers.52.self_attn.o_proj", + "model.layers.52.self_attn.gate_proj", + "model.layers.52.mlp.router.gate", + "model.layers.53.self_attn.q_proj", + "model.layers.53.self_attn.k_proj", + "model.layers.53.self_attn.v_proj", + "model.layers.53.self_attn.o_proj", + "model.layers.53.self_attn.gate_proj", + "model.layers.53.mlp.router.gate", + "model.layers.54.self_attn.q_proj", + "model.layers.54.self_attn.k_proj", + "model.layers.54.self_attn.v_proj", + "model.layers.54.self_attn.o_proj", + "model.layers.54.self_attn.gate_proj", + "model.layers.54.mlp.router.gate", + "model.layers.55.self_attn.q_proj", + "model.layers.55.self_attn.k_proj", + "model.layers.55.self_attn.v_proj", + "model.layers.55.self_attn.o_proj", + "model.layers.55.self_attn.gate_proj", + "model.layers.55.mlp.router.gate", + "model.layers.56.self_attn.q_proj", + "model.layers.56.self_attn.k_proj", + "model.layers.56.self_attn.v_proj", + "model.layers.56.self_attn.o_proj", + "model.layers.56.self_attn.gate_proj", + "model.layers.56.mlp.router.gate", + "model.layers.57.self_attn.q_proj", + "model.layers.57.self_attn.k_proj", + "model.layers.57.self_attn.v_proj", + "model.layers.57.self_attn.o_proj", + "model.layers.57.self_attn.gate_proj", + "model.layers.57.mlp.router.gate", + "model.layers.58.self_attn.q_proj", + "model.layers.58.self_attn.k_proj", + "model.layers.58.self_attn.v_proj", + "model.layers.58.self_attn.o_proj", + "model.layers.58.self_attn.gate_proj", + "model.layers.58.mlp.router.gate", + "model.layers.59.self_attn.q_proj", + "model.layers.59.self_attn.k_proj", + "model.layers.59.self_attn.v_proj", + "model.layers.59.self_attn.o_proj", + "model.layers.59.self_attn.gate_proj", + "model.layers.59.mlp.router.gate", + "lm_head" + ], + "kv_cache_scheme": null, + "quant_method": "compressed-tensors", + "quantization_status": "compressed", + "sparsity_config": {}, + "transform_config": {}, + "version": "0.14.1.a20260326" + }, + "rms_norm_eps": 1e-05, + "rope_scaling": null, + "rope_theta": 10000, + "route_norm": true, + "route_scale": 2.448, + "score_func": "sigmoid", + "sliding_window": 4096, + "tie_word_embeddings": false, + "topk_group": 1, + "transformers_version": "4.57.6", + "use_cache": true, + "use_grouped_mm": true, + "vocab_size": 200192 +} \ No newline at end of file diff --git a/configuration_afmoe.py b/configuration_afmoe.py new file mode 100644 index 0000000000000000000000000000000000000000..9efecdd517e8e6168f46ebecb3d282bdea34c5dc --- /dev/null +++ b/configuration_afmoe.py @@ -0,0 +1,133 @@ +# coding=utf-8 +# Copyright 2022 EleutherAI and the 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 transformers.configuration_utils import PretrainedConfig +from transformers.modeling_rope_utils import rope_config_validation +from transformers.configuration_utils import layer_type_validation +from transformers.utils import logging + +logger = logging.get_logger(__name__) + +class AfmoeConfig(PretrainedConfig): + """ + n_group (`int`, *optional*, defaults to 1): + Number of groups for routed experts. + topk_group (`int`, *optional*, defaults to 1): + Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups). + """ + model_type = "afmoe" + base_model_pp_plan = { + "embed_tokens": (["input_ids"], ["inputs_embeds"]), + "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), + "norm": (["hidden_states"], ["hidden_states"]), + } + + def __init__( + self, + num_hidden_layers: int = 32, + vocab_size: int = 200192, + hidden_size: int = 2048, + intermediate_size: int = 6144, + moe_intermediate_size=1408, + num_dense_layers=1, + num_attention_heads=16, + num_key_value_heads=None, + head_dim=128, + hidden_act="silu", + max_position_embeddings=16384, + initializer_range=0.02, + rms_norm_eps=1e-5, + use_cache=True, + tie_word_embeddings=False, + rope_theta=10000.0, + rope_scaling=None, + num_experts=64, + num_experts_per_tok=6, + num_shared_experts=2, + num_expert_groups=1, + num_limited_groups=1, + score_func="sigmoid", + route_norm=True, + route_scale=1.0, + global_attn_every_n_layers=4, + sliding_window=1024, + mup_enabled=False, + layer_types=None, + attention_dropout: float = 0.0, + n_group: int = 1, + topk_group: int = 1, + **kwargs, + ): + self.vocab_size = vocab_size + self.max_position_embeddings = max_position_embeddings + self.hidden_size = hidden_size + self.intermediate_size = intermediate_size + self.num_hidden_layers = num_hidden_layers + self.num_dense_layers = num_dense_layers + self.num_attention_heads = num_attention_heads + self.head_dim = head_dim + self.hidden_act = hidden_act + self.initializer_range = initializer_range + self.rms_norm_eps = rms_norm_eps + self.use_cache = use_cache + self.rope_theta = rope_theta + self.rope_scaling = rope_scaling + + + # MoE specific + self.moe_intermediate_size = moe_intermediate_size + self.num_experts_per_tok = num_experts_per_tok + self.n_group = n_group + self.topk_group = topk_group + self.num_experts = num_experts + self.num_shared_experts = num_shared_experts + self.num_expert_groups = num_expert_groups + self.num_limited_groups = num_limited_groups + self.score_func = score_func + self.route_norm = route_norm + self.route_scale = route_scale + + + # Attention specific + self.attention_dropout = attention_dropout + self.global_attn_every_n_layers = global_attn_every_n_layers + self.sliding_window = sliding_window + self.layer_types = layer_types + if self.layer_types is None: + self.layer_types = [ + "sliding_attention" if bool((i + 1) % global_attn_every_n_layers) else "full_attention" for i in range(self.num_hidden_layers) + ] + layer_type_validation(self.layer_types) + + # muP specific + self.mup_enabled = mup_enabled + + if num_key_value_heads is None: + num_key_value_heads = num_attention_heads + + self.num_key_value_heads = num_key_value_heads + + + # Validate rope configs + if self.rope_scaling is not None and "type" in self.rope_scaling: + self.rope_scaling["rope_type"] = self.rope_scaling["type"] + rope_config_validation(self) + + super().__init__( + tie_word_embeddings=tie_word_embeddings, + **kwargs, + ) + + +__all__ = ["AfmoeConfig"] diff --git a/generation_config.json b/generation_config.json new file mode 100644 index 0000000000000000000000000000000000000000..ec2172d2dddacfba7d50b745ee38f248ad989e0e --- /dev/null +++ b/generation_config.json @@ -0,0 +1,10 @@ +{ + "_from_model_config": true, + "bos_token_id": 0, + "do_sample": true, + "eos_token_id": 3, + "pad_token_id": 12, + "temperature": 0.8, + "top_p": 0.8, + "transformers_version": "4.57.6" +} diff --git a/model-00001-of-00047.safetensors b/model-00001-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..fc9671dd62d44a5b2f6da93db549dee4e8a56b3b --- /dev/null +++ b/model-00001-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:def18a0497fc8b5c53c4cb7f6d296b084cae4e405d5f3f135ffb8eb640b742ca +size 4999975928 diff --git a/model-00002-of-00047.safetensors b/model-00002-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..b888e526e22e618c69b9c49c263f58b152d16c1d --- /dev/null +++ b/model-00002-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:828ef7979cf7239a4d4e34b037bce3a2b984ae83c8ec4c9b5acd625757b721f5 +size 4996296800 diff --git a/model-00003-of-00047.safetensors b/model-00003-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7d22c55b09de22c53898ca7f03843ba51c18b6be --- /dev/null +++ b/model-00003-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1b07aeb01d86f6aef0f88ca767f28dfafea70cfce95365d25b8d73e02553558 +size 5000425320 diff --git a/model-00004-of-00047.safetensors b/model-00004-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d09d627660e821e7f73508b313e766c2658e3052 --- /dev/null +++ b/model-00004-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc41fee8c8a7e98b22b7b00ca54f5a5568d06bc32b53fbe926f6d7a855bed88a +size 4996312816 diff --git a/model-00005-of-00047.safetensors b/model-00005-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..0e5ca508c6c7c8724a8e1dc321b32c2225837db7 --- /dev/null +++ b/model-00005-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:070cbccee39c95b518352417d402d1e4156087efbb2abe1d1e1feec2e3d61186 +size 5000428416 diff --git a/model-00006-of-00047.safetensors b/model-00006-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..2c6d300cc054098ea2cf35bd2f3c0af8f315359f --- /dev/null +++ b/model-00006-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d780213a370f4b5c27c9e5628ff5764a41128b13b57c6ac5c23026fb6dc842b +size 4996300320 diff --git a/model-00007-of-00047.safetensors b/model-00007-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..efd7ed2b390d18b10e8e148c35882966e15000a9 --- /dev/null +++ b/model-00007-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b762f0e6051502b75153e1912fd07533c3fa8b8f84e98c892244b6a1a84f9b15 +size 5000429008 diff --git a/model-00008-of-00047.safetensors b/model-00008-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..240a33949c4b0f7e179dde60d788c51ee9a82b87 --- /dev/null +++ b/model-00008-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2eab2e0cee9db17c423735f2c4c3e530fadafae62708f2f86e48d547d5b5f51 +size 4996300480 diff --git a/model-00009-of-00047.safetensors b/model-00009-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..93c5a0301a4c43a217aeb82a8cef8635bca47848 --- /dev/null +++ b/model-00009-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c8e2bc499353e84d58bdef7678aef507dc29748552376f476c34968c7419e2f +size 5000444384 diff --git a/model-00010-of-00047.safetensors b/model-00010-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e80411bdaf1d68d8bce056bc8a28d314e7e8c853 --- /dev/null +++ b/model-00010-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a7a904bc2e52d5a3eb034884fa36efa60b55c08af43c8ebd27e2bd2fed2e582 +size 4996299896 diff --git a/model-00013-of-00047.safetensors b/model-00013-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d676357b8a61de400f48552e84aa74c0f18144ee --- /dev/null +++ b/model-00013-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2727bf031c2c192fa593e86ffc75438acf71e998f6c0b4feb5c760991d2dfed3 +size 5000429008 diff --git a/model-00014-of-00047.safetensors b/model-00014-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f0aac8edeefa9197541fe200154c9502cef7ad93 --- /dev/null +++ b/model-00014-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2809026dfeaf460bcbcdb2748cbcefd7279caf7f28fe049e029b6f7c0fa55544 +size 4966998400 diff --git a/model-00015-of-00047.safetensors b/model-00015-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..af1d9952bc41625d48cefd2e4d539e77f8f85d6c --- /dev/null +++ b/model-00015-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b43f5cbcd46fce8c1678ae3b448f0c5d6d2b9638019f9b64ed8af93a7af890e7 +size 4998483128 diff --git a/model-00016-of-00047.safetensors b/model-00016-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f8fc7a11ac722ba8b5c5600767d4e638b4504330 --- /dev/null +++ b/model-00016-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fe2a17f0b2bd6375dba268fc9c18c5b6749747ce6059c4f51a2fc8ceceeb1e0 +size 5000428416 diff --git a/model-00017-of-00047.safetensors b/model-00017-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1acf3ef52b863f70cf04cf4705ac5cd7eb964fbd --- /dev/null +++ b/model-00017-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a42f3d035a7733dbd1deb35522da1059076b19776cfbb3fa9c37d51f50b657c0 +size 4996300384 diff --git a/model-00018-of-00047.safetensors b/model-00018-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..732826e81f4098b3a640ecf3eab83ef3510011af --- /dev/null +++ b/model-00018-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77287564a008317747b9eed53909f2941ea88e920d79b85eb35175ddde4ecbcf +size 5000429008 diff --git a/model-00021-of-00047.safetensors b/model-00021-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..4fd4a347a839fcc939613025b1b8ae8fa733daaa --- /dev/null +++ b/model-00021-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9437db53ed6266d59e31a19d4720645cf1be87613a6ec06d9a570b262a81eb4b +size 4996299896 diff --git a/model-00022-of-00047.safetensors b/model-00022-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..2013a6dd210f6c4068c2e5bd0fdbcab16a16cb9b --- /dev/null +++ b/model-00022-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:474c665953d25b3d0e0528ace6cf5d3169ce24b9e94ba862d909f6e610e5d470 +size 5000428656 diff --git a/model-00024-of-00047.safetensors b/model-00024-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..4d12cab36c198232ccd065a2718df4077ae4351f --- /dev/null +++ b/model-00024-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7195ccbc39fecc2eda62cf0efdd454e5e5852a52f6743936ca9b837b66119dd6 +size 5000429008 diff --git a/model-00025-of-00047.safetensors b/model-00025-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5988b5334f66a05c305402fe630ae65f835b9d22 --- /dev/null +++ b/model-00025-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:480810ac10f71f9650ebc77a914dae3879e96c43bb1b6b2794fcf89fd11339ad +size 4996315968 diff --git a/model-00027-of-00047.safetensors b/model-00027-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..15dcb627d2668be1fec90a812a7fdf4f635f2ec3 --- /dev/null +++ b/model-00027-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97cced6c56e7d3247e754d3cd1ead9692ef2506141430bb358e7a9cc231bffd7 +size 4996299896 diff --git a/model-00028-of-00047.safetensors b/model-00028-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..508e9e2935a07f42461b24dbf9ded6ae6ebe65ed --- /dev/null +++ b/model-00028-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d13fbb8ae9f724b542ea8a4df8e8ff67bae14ff51cb47687590aec007ec55211 +size 5000428992 diff --git a/model-00029-of-00047.safetensors b/model-00029-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..8f3b52a04db88fd84a048a5519d6835f7c74acf9 --- /dev/null +++ b/model-00029-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33f5b9c6130b82f0a35ebf9afadd116eb5e629d9d17b2ad7e853e8d605dd908d +size 4996300480 diff --git a/model-00030-of-00047.safetensors b/model-00030-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..648eb5a689988d1c3913d3df54b232a11605415c --- /dev/null +++ b/model-00030-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:357bc4b98298e30ba38e607d0e9c2379f90727942c762cdd7991ec2dd9105fb4 +size 5000429008 diff --git a/model-00031-of-00047.safetensors b/model-00031-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..cf943dd351b13f41c76ae879786150c4b6d06496 --- /dev/null +++ b/model-00031-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9694dfa9ecb50906ff25fe3004277efbb2c77f7d02b41c04bd41c66fcfbac0e2 +size 4996315864 diff --git a/model-00032-of-00047.safetensors b/model-00032-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..b1af7d2bbafe258a80af01179bbd58dd3b4b1a47 --- /dev/null +++ b/model-00032-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a3229a2ea673e5b7edc505a81982a4fff4a955e34a65d3bea6fe0e6835fc67e +size 5000428424 diff --git a/model-00033-of-00047.safetensors b/model-00033-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..8c82893c179f58bcd83345083d173a8f889eaa40 --- /dev/null +++ b/model-00033-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b9136dfda62670c2aaa3416760ca2ed4a6e52c1850b67b69a42be25adae90a9 +size 4996300216 diff --git a/model-00035-of-00047.safetensors b/model-00035-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..51ca73d22664d4ee82a2ecb69759edfc9529614a --- /dev/null +++ b/model-00035-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:655d6363656e8bf513823155ed7f52eda9c8806ab02a96c0c146acc793340bdd +size 4996300480 diff --git a/model-00036-of-00047.safetensors b/model-00036-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..0c95a4df7b764f8d087f16532969b4c47cfc0742 --- /dev/null +++ b/model-00036-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:092ea495a451e5143e3abee0367c349ed5219f82a720d46e2c0ea3f844312bf0 +size 5000444408 diff --git a/model-00037-of-00047.safetensors b/model-00037-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a26456cb899cb7795f0965b7c93062009faf2d00 --- /dev/null +++ b/model-00037-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d5a0cf1fc975d2c537f86255e7f46f916dc6d3dfb8a68d963d1542ab6f6a667 +size 4996299880 diff --git a/model-00038-of-00047.safetensors b/model-00038-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7a29fae66f9cea0abcdf5230b1afffa4c09ec3af --- /dev/null +++ b/model-00038-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32979e95d0ed5711b18776ca5bc389b5690192645c492adef5533803fc097b10 +size 5000428496 diff --git a/model-00039-of-00047.safetensors b/model-00039-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..67cc9e95f62abdfbf3c7fc3820deb192814ccbdc --- /dev/null +++ b/model-00039-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75865fd5969e2b37649055fefe5c0385c6392a8aa552a2223615b5324adfa6b4 +size 4996300480 diff --git a/model-00040-of-00047.safetensors b/model-00040-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..46b3f86de0174e8cbfe921104076f689d6eb35b6 --- /dev/null +++ b/model-00040-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09ba6edbf10cbb4f46010d79f67fe9f82cbd70b8c6e6caa5fae9c4263725565d +size 5000429000 diff --git a/model-00041-of-00047.safetensors b/model-00041-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..b6992c7494dbee20e543771b6314851f04732255 --- /dev/null +++ b/model-00041-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91162f377e9489a4c41ac502d7fcc74892583c1b8d33bb4fe226e1c8243768b8 +size 4996300480 diff --git a/model-00042-of-00047.safetensors b/model-00042-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..6a61590f457f7752b414eb6831be952190ee3066 --- /dev/null +++ b/model-00042-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:283882498cb2b009bcca867eeb54269d859501224fc103bcb31005c06226302a +size 5000444400 diff --git a/model-00043-of-00047.safetensors b/model-00043-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..cde3c966e44b893e5d9bd0303296fa90076261e6 --- /dev/null +++ b/model-00043-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:320b965403986144b9a96894b300bea1904a2a194a3be4b7753ebef924f042d8 +size 4996299896 diff --git a/model-00044-of-00047.safetensors b/model-00044-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..df2b1cdfd9f7a166c2076423a5aabed03efedf74 --- /dev/null +++ b/model-00044-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c4116c9bbc64522bbacd6b451332df443cfa5267595bce9b2297a98fe367d7d +size 5000428832 diff --git a/model-00046-of-00047.safetensors b/model-00046-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a10aa2db28ebdd6995b443eef8612f4d03e79096 --- /dev/null +++ b/model-00046-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:356a830b736a374d7870c5a2657d03476946b80a9510d1649ae28e47310b1299 +size 5000429000 diff --git a/model-00047-of-00047.safetensors b/model-00047-of-00047.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..3362fa02777fec00650f4ea7a3b352b44ad53e86 --- /dev/null +++ b/model-00047-of-00047.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6b2daa8cf0c9de03c4bc920149c3ed5da407168c40f3e5a1138ccddbb05e5e3 +size 1612818256 diff --git a/model.safetensors.index.json b/model.safetensors.index.json new file mode 100644 index 0000000000000000000000000000000000000000..0fdb2f52f3cab121a67351c0ec4cd54365902512 --- /dev/null +++ b/model.safetensors.index.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:584b9855a0b3144a4444d2295f9a4f11d055620a3abd148db062b65b8004b800 +size 16781255 diff --git a/modeling_afmoe.py b/modeling_afmoe.py new file mode 100644 index 0000000000000000000000000000000000000000..0c22cca909b47329db43092311406bb22dd07c2c --- /dev/null +++ b/modeling_afmoe.py @@ -0,0 +1,680 @@ +from typing import Callable, Optional, Tuple, Union + +import torch +import torch.nn.functional as F +from torch import nn + +from transformers.activations import ACT2FN +from transformers.generation import GenerationMixin +from transformers.modeling_outputs import ( + MoeCausalLMOutputWithPast, + MoeModelOutputWithPast, +) +from transformers.modeling_utils import PreTrainedModel, ALL_ATTENTION_FUNCTIONS +from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update +from transformers.masking_utils import ( + create_causal_mask, + create_sliding_window_causal_mask, +) +from transformers.modeling_layers import GradientCheckpointingLayer +from transformers.processing_utils import Unpack +from transformers.utils import TransformersKwargs +from transformers.cache_utils import Cache, DynamicCache +from transformers.integrations import use_kernel_forward_from_hub + + +try: + from .configuration_afmoe import AfmoeConfig +except: + from configuration_afmoe import AfmoeConfig + +class AfmoeRotaryEmbedding(nn.Module): + + def __init__(self, config: AfmoeConfig, device=None): + super().__init__() + # BC: "rope_type" was originally "type" + if hasattr(config, "rope_scaling") and config.rope_scaling is not None: + self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.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 + + 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 + + def _dynamic_frequency_update(self, position_ids, device): + """ + dynamic RoPE layers should recompute `inv_freq` in the following situations: + 1 - growing beyond the cached sequence length (allow scaling) + 2 - the current sequence length is in the original scale (avoid losing precision with small sequences) + """ + seq_len = torch.max(position_ids) + 1 + if seq_len > self.max_seq_len_cached: # growth + inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len) + self.register_buffer("inv_freq", inv_freq, persistent=False) # TODO joao: may break with compilation + self.max_seq_len_cached = seq_len + + if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len: # reset + # This .to() is needed if the model has been moved to a device after being initialized (because + # the buffer is automatically moved, but not the original copy) + self.original_inv_freq = self.original_inv_freq.to(device) + self.register_buffer("inv_freq", self.original_inv_freq, persistent=False) + self.max_seq_len_cached = self.original_max_seq_len + + @torch.no_grad() + def forward(self, x, position_ids): + if "dynamic" in self.rope_type: + self._dynamic_frequency_update(position_ids, device=x.device) + + # Core RoPE block + inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) + position_ids_expanded = position_ids[:, None, :].float() + # Force float32 (see https://github.com/huggingface/transformers/pull/29285) + device_type = x.device.type + device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu" + with torch.autocast(device_type=device_type, enabled=False): + freqs = (inv_freq_expanded.float().to(x.device) @ position_ids_expanded.float()).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + cos = emb.cos() + sin = emb.sin() + + # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention + cos = cos * self.attention_scaling + sin = 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. + """ + cos = cos.unsqueeze(unsqueeze_dim) + sin = sin.unsqueeze(unsqueeze_dim) + q_embed = (q * cos) + (rotate_half(q) * sin) + k_embed = (k * cos) + (rotate_half(k) * sin) + 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) + +@use_kernel_forward_from_hub("RMSNorm") +class AfmoeRMSNorm(nn.Module): + def __init__(self, hidden_size: int, eps: float): + """ + AfmoeRMSNorm is equivalent to T5LayerNorm + """ + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + hidden_states = hidden_states.to(torch.float32) + variance = hidden_states.pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + return self.weight * hidden_states.to(input_dtype) + + def extra_repr(self): + return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" + + + +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) + + 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, dtype=torch.float32).to( + query.dtype + ) + 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 + + +class AfmoeMLP(nn.Module): + def __init__(self, config, intermediate_size=None): + super().__init__() + self.config = config + self.hidden_size = config.hidden_size + self.intermediate_size = intermediate_size or 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[config.hidden_act] + + def forward(self, x): + return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) + + +class AfmoeTokenChoiceRouter(nn.Module): + """Token-choice top-K router for MoE routing.""" + + def __init__(self, config): + super().__init__() + self.config = config + self.top_k = config.num_experts_per_tok + self.num_experts = config.num_experts + self.score_func = config.score_func + self.route_norm = config.route_norm + self.route_scale = config.route_scale + self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False) + + def forward(self, hidden_states, expert_bias: torch.Tensor | None): + _, _, hidden_dim = hidden_states.shape + hidden_states = hidden_states.view(-1, hidden_dim) + + scores = self.gate(hidden_states) + + # Apply scoring function in float32 for stability + if self.score_func == "sigmoid": + scores = torch.sigmoid(scores.to(torch.float32)) + else: + scores = F.softmax(scores.to(torch.float32), dim=-1) + + if expert_bias is not None: + _, selected_experts = torch.topk(scores + expert_bias, k=self.top_k, dim=1) + top_scores = scores.gather(dim=1, index=selected_experts) + else: + top_scores, selected_experts = torch.topk(scores, k=self.top_k, dim=1) + + # Normalize weights if using sigmoid + if self.score_func == "sigmoid" and self.route_norm: + denominator = top_scores.sum(dim=-1, keepdim=True) + 1e-20 + top_scores = top_scores / denominator + + top_scores = top_scores * self.route_scale + return top_scores, selected_experts + +class AfmoeMoE(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.router = AfmoeTokenChoiceRouter(config) + + self.shared_experts = None + if config.num_shared_experts > 0: + self.shared_experts = AfmoeMLP( + config, config.moe_intermediate_size * config.num_shared_experts + ) + self.experts = nn.ModuleList( + [AfmoeMLP( + config, intermediate_size=config.moe_intermediate_size + ) for _ in range(config.num_experts)] + ) + self.expert_bias = nn.Parameter(torch.zeros(config.num_experts, dtype=torch.float32), requires_grad=False) + + + def forward(self, hidden_states): + batch_size, seq_len, hidden_dim = hidden_states.shape + hidden_states_flat = hidden_states.view(-1, hidden_dim) + + # Get routing decisions + top_scores, selected_experts = self.router(hidden_states, self.expert_bias) + + # Process through shared experts + if self.shared_experts is not None: + shared_output = self.shared_experts(hidden_states_flat) + else: + shared_output = torch.zeros_like(hidden_states_flat) + + # Reorder tokens by expert for efficient processing + token_indices_sorted = torch.argsort(selected_experts.view(-1), stable=True) + top_scores_sorted = top_scores.view(-1)[token_indices_sorted] + token_to_expert = selected_experts.view(-1)[token_indices_sorted] + token_indices_sorted = token_indices_sorted // self.config.num_experts_per_tok + + # Gather input tokens + token_indices_expanded = token_indices_sorted.unsqueeze(-1).expand( + -1, hidden_dim + ) + routed_input = torch.gather( + hidden_states_flat, dim=0, index=token_indices_expanded + ) + + routed_output = torch.zeros_like(routed_input) + for expert_id in range(self.config.num_experts): + mask = token_to_expert == expert_id + if mask.any(): + expert_input = routed_input[mask] + expert_out = self.experts[expert_id](expert_input) + routed_output[mask] = expert_out + + routed_output = ( + routed_output.to(torch.float32) * top_scores_sorted.unsqueeze(-1) + ).to(hidden_states.dtype) + + # Scatter back to original positions + output = shared_output.scatter_add( + dim=0, index=token_indices_expanded, src=routed_output + ) + + return output.view(batch_size, seq_len, hidden_dim) + + +class AfmoeAttention(nn.Module): + """Multi-headed attention with local/global pattern and gating.""" + + def __init__(self, config: AfmoeConfig, layer_idx: int): + super().__init__() + self.config = config + self.layer_idx = layer_idx + self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads) + self.num_heads = config.num_attention_heads + self.num_key_value_heads = config.num_key_value_heads + self.num_key_value_groups = self.num_heads // self.num_key_value_heads + + self.scaling = self.head_dim**-0.5 + self.attention_dropout = config.attention_dropout + self.is_local_attention = config.layer_types[layer_idx] == "sliding_attention" + self.sliding_window = config.sliding_window if self.is_local_attention else None + + self.q_proj = nn.Linear( + config.hidden_size, self.num_heads * self.head_dim, bias=False + ) + self.k_proj = nn.Linear( + config.hidden_size, self.num_key_value_heads * self.head_dim, bias=False + ) + self.v_proj = nn.Linear( + config.hidden_size, self.num_key_value_heads * self.head_dim, bias=False + ) + self.o_proj = nn.Linear( + self.num_heads * self.head_dim, config.hidden_size, bias=False + ) + + self.q_norm = AfmoeRMSNorm(self.head_dim, eps=config.rms_norm_eps) + self.k_norm = AfmoeRMSNorm(self.head_dim, eps=config.rms_norm_eps) + + self.gate_proj = nn.Linear( + config.hidden_size, self.num_heads * self.head_dim, bias=False + ) + + def forward( + self, + hidden_states: torch.Tensor, + position_embeddings: tuple[torch.Tensor, torch.Tensor], + attention_mask: Optional[torch.Tensor], + past_key_value: Optional[Cache] = None, + cache_position: Optional[torch.LongTensor] = None, + **kwargs: Unpack[TransformersKwargs], + ) -> torch.Tensor: + + input_shape = hidden_states.shape[:-1] + hidden_shape = (*input_shape, -1, self.head_dim) + + query_states = self.q_proj(hidden_states).view(hidden_shape) + key_states = self.k_proj(hidden_states).view(hidden_shape) + value_states = self.v_proj(hidden_states).view(hidden_shape) + gate_states = self.gate_proj(hidden_states) + + query_states = self.q_norm(query_states) + key_states = self.k_norm(key_states) + + query_states = query_states.transpose(1, 2) + key_states = key_states.transpose(1, 2) + value_states = value_states.transpose(1, 2) + + if self.is_local_attention: + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + if past_key_value is not None: + cache_kwargs = {"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 + if self.config._attn_implementation != "eager": + attention_interface = ALL_ATTENTION_FUNCTIONS[ + self.config._attn_implementation + ] + + output, _ = attention_interface( + self, + query_states, + key_states, + value_states, + attention_mask=attention_mask, + dropout=0.0 if not self.training else self.attention_dropout, + scaling=self.scaling, + sliding_window=self.sliding_window, + **kwargs, + ) + + output = output.view(*input_shape, -1).contiguous() + output = output * F.sigmoid(gate_states) + return self.o_proj(output) + + +class AfmoeDecoderLayer(GradientCheckpointingLayer): + def __init__(self, config: AfmoeConfig, layer_idx: int): + super().__init__() + self.hidden_size = config.hidden_size + self.layer_idx = layer_idx + + self.self_attn = AfmoeAttention(config=config, layer_idx=layer_idx) + self.attention_type = config.layer_types[layer_idx] + + # Dual normalization for attention + self.input_layernorm = AfmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = AfmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + # Dual normalization for FFN + self.pre_mlp_layernorm = AfmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_mlp_layernorm = AfmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + # MoE or dense FFN + self.moe_enabled = layer_idx >= config.num_dense_layers + if self.moe_enabled: + self.mlp = AfmoeMoE(config) + else: + self.mlp = AfmoeMLP(config) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + use_cache: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None, + **kwargs: Unpack[TransformersKwargs], + ) -> torch.FloatTensor: + residual = hidden_states + + # Self Attention with dual normalization + 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, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **kwargs, + ) + hidden_states = self.post_attention_layernorm(hidden_states) + hidden_states = residual + hidden_states + + # FFN with dual normalization + residual = hidden_states + hidden_states = self.pre_mlp_layernorm(hidden_states) + + if self.moe_enabled: + hidden_states = self.mlp(hidden_states) + else: + hidden_states = self.mlp(hidden_states) + + hidden_states = self.post_mlp_layernorm(hidden_states) + hidden_states = residual + hidden_states + return hidden_states + + +class AfmoePreTrainedModel(PreTrainedModel): + config_class = AfmoeConfig + base_model_prefix = "model" + _no_split_modules = ["AfmoeDecoderLayer"] + _skip_keys_device_placement = ["past_key_values"] + _keep_in_fp32_modules = [ + "input_layernorm", + "post_attention_layernorm", + "pre_mlp_layernorm", + "post_mlp_layernorm", + "q_norm", + "k_norm", + "norm", + ] + _supports_sdpa = True + _supports_attention_backend = True + supports_gradient_checkpointing = True + + +class AfmoeModel(AfmoePreTrainedModel): + _no_split_modules = ["AfmoeDecoderLayer"] + + def __init__(self, config: AfmoeConfig): + 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( + [ + AfmoeDecoderLayer(config, layer_idx) + for layer_idx in range(config.num_hidden_layers) + ] + ) + self.norm = AfmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.rotary_emb = AfmoeRotaryEmbedding(config=config) + self.gradient_checkpointing = False + + self.post_init() + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + + def forward( + self, + input_ids: torch.LongTensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[list[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + **kwargs: Unpack[TransformersKwargs], + ) -> MoeModelOutputWithPast: + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You must specify exactly one of input_ids or inputs_embeds" + ) + + if use_cache and past_key_values is None: + past_key_values = DynamicCache() + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + 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) + + # It may already have been prepared by e.g. `generate` + if not isinstance(causal_mask_mapping := attention_mask, dict): + mask_kwargs = { + "config": self.config, + "input_embeds": inputs_embeds, + "attention_mask": attention_mask, + "cache_position": cache_position, + "past_key_values": past_key_values, + } + causal_mask_mapping = { + "full_attention": create_causal_mask(**mask_kwargs), + "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs), + } + + hidden_states = inputs_embeds + + # Apply muP input scaling if enabled + if self.config.mup_enabled: + hidden_states = hidden_states * (self.config.hidden_size**0.5) + + position_embeddings = self.rotary_emb(hidden_states, position_ids) + + for decoder_layer in self.layers: + hidden_states = decoder_layer( + hidden_states, + attention_mask=causal_mask_mapping[decoder_layer.attention_type], + position_ids=position_ids, + past_key_value=past_key_values, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **kwargs, + ) + + hidden_states = self.norm(hidden_states) + return MoeModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=past_key_values, + ) + + +class AfmoeForCausalLM(AfmoePreTrainedModel, GenerationMixin): + _tied_weights_keys = ["lm_head.weight"] + _tp_plan = {"lm_head": "colwise_rep"} + _pp_plan = {"lm_head": (["hidden_states"], ["logits"])} + + def __init__(self, config): + super().__init__(config) + self.model = AfmoeModel(config) + self.vocab_size = config.vocab_size + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def set_decoder(self, decoder): + self.model = decoder + + def get_decoder(self): + return self.model + + def forward( + self, + input_ids: torch.LongTensor, + 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, + cache_position: Optional[torch.LongTensor] = None, + logits_to_keep: Union[int, torch.Tensor] = 0, + token_type_ids: Optional[torch.Tensor] = None, # will be ignored + **kwargs: Unpack[TransformersKwargs], + ) -> Union[Tuple, MoeCausalLMOutputWithPast]: + outputs: MoeModelOutputWithPast = self.model( + input_ids=input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + cache_position=cache_position, + **kwargs, + ) + + hidden_states = outputs.last_hidden_state + # Only compute necessary logits + slice_indices = ( + slice(-logits_to_keep, None) + if isinstance(logits_to_keep, int) + else logits_to_keep + ) + logits = self.lm_head(hidden_states[:, slice_indices, :]) + + loss = None + if labels is not None: + loss = self.loss_function(logits, labels, self.vocab_size, **kwargs) + + + return MoeCausalLMOutputWithPast( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + router_logits=outputs.router_logits, + ) + + +__all__ = [ + "AfmoeForCausalLM", + "AfmoeModel", + "AfmoePreTrainedModel", +] diff --git a/recipe.yaml b/recipe.yaml new file mode 100644 index 0000000000000000000000000000000000000000..c3887f15dcd9dab723d7c8a9cb7ea29739021857 --- /dev/null +++ b/recipe.yaml @@ -0,0 +1,7 @@ +default_stage: + default_modifiers: + QuantizationModifier: + targets: [Linear] + ignore: [lm_head, 're:.*self_attn.*', 're:.*mlp.router.*'] + scheme: NVFP4 + bypass_divisibility_checks: false diff --git a/special_tokens_map.json b/special_tokens_map.json new file mode 100644 index 0000000000000000000000000000000000000000..5fd26c542275c864fff98316bef46b902bc68cce --- /dev/null +++ b/special_tokens_map.json @@ -0,0 +1,23 @@ +{ + "bos_token": { + "content": "<|begin_of_text|>", + "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": { + "content": "<|pad|>", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false + } +} diff --git a/tokenizer.json b/tokenizer.json new file mode 100644 index 0000000000000000000000000000000000000000..20286e2b0432a2f82d206ea4a2788e51775e376e --- /dev/null +++ b/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5a93d847b4d3a1da95e9527c30ec10144f63a823e9feec98570274980754897 +size 14614487 diff --git a/tokenizer_config.json b/tokenizer_config.json new file mode 100644 index 0000000000000000000000000000000000000000..88320d663cb9cbb25a9f5348166bee5af8209901 --- /dev/null +++ b/tokenizer_config.json @@ -0,0 +1,271 @@ +{ + "add_bos_token": false, + "add_eos_token": false, + "add_prefix_space": null, + "added_tokens_decoder": { + "0": { + "content": "<|begin_of_text|>", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "1": { + "content": "<|end_of_text|>", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "2": { + "content": "<|im_start|>", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "3": { + "content": "<|im_end|>", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "4": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "5": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "6": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "7": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "8": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "9": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "10": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "11": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "12": { + "content": "<|pad|>", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + }, + "13": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "14": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "15": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "16": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "17": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "18": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "19": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "20": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "21": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "22": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "23": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "24": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "25": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "26": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "27": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "28": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": false + }, + "29": { + "content": "", + "lstrip": false, + "normalized": false, + "rstrip": false, + "single_word": false, + "special": true + } + }, + "bos_token": "<|begin_of_text|>", + "clean_up_tokenization_spaces": false, + "eos_token": "<|im_end|>", + "extra_special_tokens": {}, + "model_max_length": 65536, + "pad_token": "<|pad|>", + "tokenizer_class": "PreTrainedTokenizerFast", + "use_default_system_prompt": false +}