IMPORTANT: To view this page as Markdown, append `.md` to the URL (e.g. /max/get-started.md). For the complete documentation index, see llms.txt.
Skip to main content
For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Python module

max.pipelines.architectures.kimik2_5

Kimi K2.5 mixture-of-experts architecture for text generation.

KimiK2_5Config​

class max.pipelines.architectures.kimik2_5.KimiK2_5Config(*, devices, dtype, bos_token_id, eos_token_id, ignore_index, media_placeholder_token_id, pad_token_id, tie_word_embeddings, use_unified_vision_chunk, video_placeholder, vision_config, llm_config)

source

Bases: ArchVLConfigWithTextSubconfig, ArchConfigWithKVCache

Configuration for Kimi-K2.5 models.

Parameters:

bos_token_id​

bos_token_id: int

source

ID of the beginning-of-sequence (BOS) token.

devices​

devices: list[DeviceRef]

source

Devices that the Kimi-K2.5 model is parallelized over.

dtype​

dtype: DType

source

DType of the Kimi-K2.5 model weights.

eos_token_id​

eos_token_id: int

source

ID of the end-of-sequence (EOS) token.

get_kv_params()​

get_kv_params()

source

Returns the KV cache parameters from the embedded LLM config.

Return type:

KVCacheParamInterface

get_num_layers()​

static get_num_layers(huggingface_config)

source

Parameters:

huggingface_config (AutoConfig)

Return type:

int

ignore_index​

ignore_index: int

source

Index that should be ignored when calculating loss (e.g., for padding).

initialize()​

classmethod initialize(pipeline_config, model_config=None)

source

Initializes a Qwen3VLConfig instance from pipeline configuration.

Parameters:

Returns:

A Qwen3VLConfig instance with fields initialized from config.

Return type:

Self

initialize_from_config()​

classmethod initialize_from_config(pipeline_config, huggingface_config, llm_config=None)

source

Initializes a KimiK2_5Config from pipeline and HuggingFace configs.

This method creates a config instance with all fields that can be determined from the pipeline and HuggingFace configurations, without needing the state_dict. Fields that depend on the state_dict should be set via the finalize() method.

Parameters:

  • pipeline_config (PipelineConfig) – The MAX Engine pipeline configuration.
  • huggingface_config (AutoConfig) – HuggingFace model configuration.
  • llm_config (KimiK2_5TextConfig | None) – Pre-initialized DeepseekV3 configuration.

Returns:

A KimiK2_5Config instance ready for finalization.

Return type:

Self

llm_config​

llm_config: KimiK2_5TextConfig

source

Language model configuration using DeepseekV3 architecture.

media_placeholder_token_id​

media_placeholder_token_id: int

source

Token ID used as a placeholder for media (e.g., images, video frames) within sequences.

pad_token_id​

pad_token_id: int

source

Token ID used for padding sequences to uniform length.

tie_word_embeddings​

tie_word_embeddings: bool

source

Whether to share (tie) the input and output word embeddings in the language model.

use_unified_vision_chunk​

use_unified_vision_chunk: bool | None

source

Whether to use a unified chunk for vision inputs.

video_placeholder​

video_placeholder: str | None

source

Placeholder string used to represent video segments in input text.

vision_config​

vision_config: VisionConfig

source

Vision encoder configuration.

KimiK2_5Model​

class max.pipelines.architectures.kimik2_5.KimiK2_5Model(pipeline_config, session, devices, kv_cache_config, weights, adapter=None, return_logits=ReturnLogits.ALL, return_hidden_states=ReturnHiddenStates.NONE, max_batch_size=1)

source

Bases: AlwaysSignalBuffersMixin, MultiGraphPipelineModelWithKVCache[KimiK2_5TextAndVisionContext]

A Kimi-K2.5 pipeline model for multimodal text generation.

Parameters:

batch_processor_cls​

batch_processor_cls

source

alias of KimiK2_5BatchProcessor

execute()​

execute(model_inputs)

source

Executes the graph with the given inputs.

Parameters:

model_inputs (ModelInputs) – The model inputs to execute, containing tensors and any other required data for model execution.

Returns:

ModelOutputs containing the pipeline’s output tensors.

Return type:

ModelOutputs

This is an abstract method that must be implemented by concrete PipelineModels to define their specific execution logic.

get_kv_params()​

classmethod get_kv_params(huggingface_config, pipeline_config, devices, kv_cache_config, cache_dtype)

source

Returns the KV cache params for the pipeline model.

Delegates to model_config_cls.construct_kv_params(...). Subclasses with custom KV behavior should override this method.

Parameters:

Return type:

KVCacheParamInterface

language_model​

language_model: Model

source

The compiled language model for text generation.

load_model()​

load_model(session)

source

Build, compile, and load vision and language graphs into session.

Parameters:

session (InferenceSession)

Return type:

tuple[Model | None, Model]

model​

property model: Model

source

Expose language model for graph capture/replay.

Only the language model is captured since vision runs during prefill

model_config_cls​

model_config_cls

source

alias of KimiK2_5Config

prepare_initial_token_inputs()​

prepare_initial_token_inputs(replica_batches, kv_cache_inputs=None, return_n_logits=1)

source

Delegates to the batch processor; typed for Eagle subclasses.

Parameters:

Return type:

KimiK2_5ModelInputs

release()​

release(request_id)

source

Release vision encoder cache entries for a completed request.

Parameters:

request_id (RequestID)

Return type:

None

vision_model​

vision_model: Model | None

source

The compiled vision model for processing images.

KimiK2_5ModelInputs​

class max.pipelines.architectures.kimik2_5.KimiK2_5ModelInputs(tokens, input_row_offsets, signal_buffers, host_input_row_offsets, batch_context_lengths, image_token_indices=None, precomputed_image_embeddings=None, pixel_values=None, grid_thws=None, cu_seqlens=None, max_seqlen=None, vision_position_ids=None, language_image_embeddings=<factory>, language_image_token_indices=<factory>, eplb_counter_buffers=<factory>, *, kv_cache_inputs=None, lora=None, hidden_states=None, return_n_logits, data_parallel_splits, ep_inputs=())

source

Bases: DeepseekV3Inputs

A class representing inputs for the KimiK2_5M model.

This class encapsulates the input tensors required for the KimiK2_5M model execution, including both text and vision inputs. Vision inputs are optional and can be None for text-only processing.

Parameters:

buffers​

property buffers: tuple[Buffer, ...]

source

Returns the language model input ABI tuple.

cu_seqlens​

cu_seqlens: list[Buffer] | None = None

source

Cumulative sequence lengths for full attention per device.

eplb_counter_buffers​

eplb_counter_buffers: list[Buffer]

source

Per-device EP counter buffers for the language model graph.

grid_thws​

grid_thws: list[Buffer] | None = None

source

Grid dimensions (temporal, height, width) for each image/video, shape (n_images, 3) per device.

has_vision_inputs​

property has_vision_inputs: bool

source

Check if this input contains vision data.

image_token_indices​

image_token_indices: list[Buffer] | None = None

source

Per-device pre-computed multimodal merge indices for the image embeddings.

These are the locations of the image_token_id in the inputs fed to the model.

Some indices may be negative, which means that they are ignored by the multimodal merge.

language_image_embeddings​

language_image_embeddings: list[Buffer]

source

Per-device image embeddings for the language model graph. Shape [0, hidden_size] during decode, [num_patches, hidden_size] during prefill.

language_image_token_indices​

language_image_token_indices: list[Buffer]

source

Per-device scatter indices for the language model graph. Shape [0] during decode, [num_image_tokens] during prefill.

max_seqlen​

max_seqlen: list[Buffer] | None = None

source

Maximum sequence length for full attention for vision inputs per device.

pixel_values​

pixel_values: list[Buffer] | None = None

source

Pixel values for vision inputs.

precomputed_image_embeddings​

precomputed_image_embeddings: list[Buffer] | None = None

source

Pre-computed image embeddings from VisionEncoderCache.

vision_position_ids​

vision_position_ids: list[Buffer] | None = None

source

Vision rotary position IDs per device.

KimiK2_5ReasoningParser​

class max.pipelines.architectures.kimik2_5.KimiK2_5ReasoningParser(think_start_token_id, think_end_token_id, tool_section_start_token_id=None)

source

Bases: ReasoningParser

Kimi K2.5 reasoning parser for … sections.

Per Moonshot’s “Interleaved Thinking” design (see https://platform.moonshot.ai/docs/guide/use-kimi-k2-thinking-model and https://huggingface.co/moonshotai/Kimi-K2.5), a single assistant turn can interleave multiple <think>...</think> blocks with <|tool_calls_section_begin|>...<|tool_calls_section_end|> blocks.

A reasoning span ends on </think> or <|tool_calls_section_begin|>; the model may open the tool-call section directly from inside the prefilled <think> block without a closing </think>. The section marker is left as content rather than consumed as a delimiter, so the tool parser (which only sees content) receives the whole section.

Reasoning may begin implicitly, without an explicit <think> token, when the chat template prefilled the assistant turn already inside a thinking block.

Reasoning can be disabled through the chat template by including a </think> token at the end of the prompt; this is detected by will_reason_after_prompt().

Parameters:

  • think_start_token_id (int)
  • think_end_token_id (int)
  • tool_section_start_token_id (int | None)

from_tokenizer()​

async classmethod from_tokenizer(tokenizer)

source

Construct a reasoning parser from a tokenizer.

Parameters:

tokenizer (PipelineTokenizer[Any, Any, Any])

Return type:

KimiK2_5ReasoningParser

reasoning_end_token_id()​

async classmethod reasoning_end_token_id(tokenizer)

source

Returns the </think> token id.

Parameters:

tokenizer (PipelineTokenizer[Any, Any, Any])

Return type:

int | None

stream()​

stream(delta_token_ids, is_currently_reasoning=True)

source

Identify a reasoning span within a streaming delta chunk.

When is_currently_reasoning=False and the chunk contains no <think> opener, returns an empty span so non-reasoning chunks (turns where the chat template prefilled </think>, or any chunk after reasoning ended in a prior chunk) aren’t misclassified as reasoning.

Parameters:

Return type:

ParsedReasoningDelta

will_reason_after_prompt()​

will_reason_after_prompt(prompt_token_ids)

source

Predicts whether the model will emit reasoning after this prompt.

Kimi K2.5 chat templates emit <think> to open the new assistant turn’s reasoning section, and </think> to close the prior assistant turn’s reasoning section.

Scan right-to-left and return based on the first delimiter seen:

  • <think> → reasoning is currently open → True.
  • </think> (or <|tool_calls_section_begin|>) → reasoning is currently closed → False.
  • No delimiters at all → reasoning is not in use → False.

Uses the same end-of-reasoning delimiters as stream() so both agree on where reasoning ends.

Parameters:

prompt_token_ids (Sequence[int])

Return type:

bool

KimiK2_5TextConfig​

class max.pipelines.architectures.kimik2_5.KimiK2_5TextConfig(*, dtype: 'DType', kv_params: 'KVCacheParamInterface', devices: 'list[DeviceRef]', use_subgraphs: 'bool' = True, data_parallel_degree: 'int' = 1, vocab_size: 'int' = 129280, hidden_size: 'int' = 7168, intermediate_size: 'int' = 18432, moe_intermediate_size: 'int' = 2048, moe_layer_freq: 'int' = 1, num_hidden_layers: 'int' = 61, num_attention_heads: 'int' = 128, num_key_value_heads: 'int' = 128, n_shared_experts: 'int' = 1, n_routed_experts: 'int' = 256, routed_scaling_factor: 'float' = 2.5, kv_lora_rank: 'int' = 512, q_lora_rank: 'int' = 1536, qk_rope_head_dim: 'int' = 64, v_head_dim: 'int' = 128, qk_nope_head_dim: 'int' = 128, topk_method: 'str' = 'greedy', n_group: 'int' = 8, topk_group: 'int' = 4, num_experts_per_tok: 'int' = 8, first_k_dense_replace: 'int' = 3, norm_topk_prob: 'bool' = True, hidden_act: 'str' = 'silu', max_position_embeddings: 'int' = 4096, max_seq_len: 'int' = 163840, rms_norm_eps: 'float' = 1e-06, tie_word_embeddings: 'bool' = False, rope_theta: 'float' = 10000.0, rope_scaling: 'dict[str, Any] | None' = None, rope_interleave: 'bool' = True, scoring_func: 'str' = 'sigmoid', attention_bias: 'bool' = False, attention_dropout: 'float' = 0.0, norm_dtype: 'DType' = bfloat16, gate_dtype: 'DType | None' = None, correction_bias_dtype: 'DType | None' = None, max_batch_context_length: 'int' = 131072, quant_config: 'QuantConfig | None' = None, dense_mlp_layers_without_quant: 'frozenset[int]' = frozenset(), ep_config: 'EPConfig | None' = None, graph_mode: 'str' = 'auto', return_logits: 'ReturnLogits' = <ReturnLogits.LAST_TOKEN: 'last_token'>, return_hidden_states: 'ReturnHiddenStates' = <ReturnHiddenStates.NONE: 'none'>, eagle_aux_hidden_state_layer_ids: 'list[int] | None' = None, eplb_profile_enabled: 'bool' = False)

source

Bases: DeepseekV3Config

Parameters:

  • dtype (DType)
  • kv_params (KVCacheParamInterface)
  • devices (list[DeviceRef])
  • use_subgraphs (bool)
  • data_parallel_degree (int)
  • vocab_size (int)
  • hidden_size (int)
  • intermediate_size (int)
  • moe_intermediate_size (int)
  • moe_layer_freq (int)
  • num_hidden_layers (int)
  • num_attention_heads (int)
  • num_key_value_heads (int)
  • n_shared_experts (int)
  • n_routed_experts (int)
  • routed_scaling_factor (float)
  • kv_lora_rank (int)
  • q_lora_rank (int)
  • qk_rope_head_dim (int)
  • v_head_dim (int)
  • qk_nope_head_dim (int)
  • topk_method (str)
  • n_group (int)
  • topk_group (int)
  • num_experts_per_tok (int)
  • first_k_dense_replace (int)
  • norm_topk_prob (bool)
  • hidden_act (str)
  • max_position_embeddings (int)
  • max_seq_len (int)
  • rms_norm_eps (float)
  • tie_word_embeddings (bool)
  • rope_theta (float)
  • rope_scaling (dict[str, Any] | None)
  • rope_interleave (bool)
  • scoring_func (str)
  • attention_bias (bool)
  • attention_dropout (float)
  • norm_dtype (DType)
  • gate_dtype (DType | None)
  • correction_bias_dtype (DType | None)
  • max_batch_context_length (int)
  • quant_config (QuantConfig | None)
  • dense_mlp_layers_without_quant (frozenset[int])
  • ep_config (EPConfig | None)
  • graph_mode (str)
  • return_logits (ReturnLogits)
  • return_hidden_states (ReturnHiddenStates)
  • eagle_aux_hidden_state_layer_ids (list[int] | None)
  • eplb_profile_enabled (bool)

calculate_max_seq_len()​

classmethod calculate_max_seq_len(pipeline_config, huggingface_config, model_config=None)

source

Parameters:

Return type:

int

initialize()​

classmethod initialize(pipeline_config, model_config=None)

source

Initializes a DeepseekV3Config instance from pipeline configuration.

This method creates a config instance with all fields that can be determined from the pipeline configuration, without needing the state_dict. Fields that depend on the state_dict (like norm_dtype, quant_config, etc.) should be set via the finalize() method.

Parameters:

Returns:

An initialized DeepseekV3Config instance.

Return type:

Self

KimiToolParser​

class max.pipelines.architectures.kimik2_5.KimiToolParser

source

Bases: StructuralTagToolParser

Parses Kimi K2.5-style tool calls from model responses.

Kimi K2.5 wraps tool calls in section/call markers and embeds the function name as a compound functions.{name}:{idx} identifier before a dedicated argument-begin marker. Arguments are raw JSON, which the base class can diff directly.

CALL_BEGIN​

CALL_BEGIN: ClassVar[str] = '<|tool_call_begin|>'

source

CALL_END​

CALL_END: ClassVar[str] = '<|tool_call_end|>'

source

SECTION_BEGIN​

SECTION_BEGIN: ClassVar[str] = '<|tool_calls_section_begin|>'

source

SECTION_END​

SECTION_END: ClassVar[str] = '<|tool_calls_section_end|>'

source

XGRAMMAR_FORMAT​

XGRAMMAR_FORMAT = 'kimi'

source

generate_tool_call_grammar()​

static generate_tool_call_grammar(response_format_schema=None, tools=None, tokenizer=None, backend='xgrammar', tool_choice=None, **kwargs)

source

Generates a grammar for constrained decoding of Kimi tool calls.

With the default backend="xgrammar" this returns a serialized xgrammar StructuralTag (which constrains each call’s arguments to that tool’s JSON schema). With backend="llguidance" it returns a Lark grammar whose argument body is freeform.

Kimi K2.5 performs “interleaved thinking”: a single assistant turn can interleave multiple <think>...</think> reasoning blocks with multiple <|tool_calls_section_begin|>...<|tool_calls_section_end|> tool-call sections, and ends the turn with <|im_end|>. The grammar admits up to _MAX_TOOL_CALL_SECTIONS sections, an optional reasoning block before each, and an optional trailing <|im_end|> so the model can stop before the cap.

Structural markers, <think>/</think>, and <|im_end|> are referenced as single-token symbols (<[id]>) resolved from tokenizer — they are atomic special tokens, so the freeform /[\s\S]*/ argument and reasoning bodies terminate cleanly at the closing marker. Reasoning enforced this way is plain text; a mid-reasoning special token is not admitted under forced decoding.

When response_format_schema is provided, the grammar also accepts a JSON response matching the schema (the model’s first tokens select the branch).

Parameters:

  • response_format_schema (dict[str, Any] | None) – Optional JSON schema dict. When provided, the grammar also accepts a JSON response matching the schema.
  • tools (list[dict[str, Any]] | None) – Optional list of OpenAI-style tool dicts. None accepts any length-capped identifier as the function name.
  • tokenizer (PipelineTokenizer[Any, Any, Any] | None) – Pipeline tokenizer used to resolve special-token IDs. Required.
  • **kwargs (Any) – Ignored; accepts future kwargs.
  • backend (str)
  • tool_choice (str | dict[str, Any] | None)
  • **kwargs

Returns:

A grammar string compatible with the selected backend.

Return type:

str

VisionConfig​

class max.pipelines.architectures.kimik2_5.VisionConfig(dtype, devices, init_pos_emb_height, init_pos_emb_time, init_pos_emb_width, merge_kernel_size, mm_hidden_size, patch_size, projector_ln_eps, text_hidden_size, vt_hidden_size, vt_intermediate_size, vt_num_attention_heads, vt_num_hidden_layers, merge_type=None, mm_projector_type=None, model_type='', pos_emb_type=None, projector_hidden_act=None, video_attn_type=None, has_bias=True, in_channels=3, rope_max_height=512, rope_max_width=512, rope_theta=10000.0)

source

Bases: object

Vision configuration for Kimi-K2.5 models with required fields.

Parameters:

  • dtype (DType)
  • devices (list[DeviceRef])
  • init_pos_emb_height (int)
  • init_pos_emb_time (int)
  • init_pos_emb_width (int)
  • merge_kernel_size (list[int])
  • mm_hidden_size (int)
  • patch_size (int)
  • projector_ln_eps (float)
  • text_hidden_size (int)
  • vt_hidden_size (int)
  • vt_intermediate_size (int)
  • vt_num_attention_heads (int)
  • vt_num_hidden_layers (int)
  • merge_type (str | None)
  • mm_projector_type (str | None)
  • model_type (str)
  • pos_emb_type (str | None)
  • projector_hidden_act (str | None)
  • video_attn_type (str | None)
  • has_bias (bool)
  • in_channels (int)
  • rope_max_height (int)
  • rope_max_width (int)
  • rope_theta (float)

devices​

devices: list[DeviceRef]

source

Devices that the Kimi-K2.5 vision encoder model is parallelized over.

dtype​

dtype: DType

source

DType of the Kimi-K2.5 vision model weights.

finalize()​

finalize(vision_dtype)

source

Finalize VisionConfig with state_dict dependent fields.

Parameters:

vision_dtype (DType)

Return type:

None

has_bias​

has_bias: bool = True

source

Whether linear projections in the vision transformer include bias terms.

in_channels​

in_channels: int = 3

source

Number of input image channels (3 for RGB).

init_pos_emb_height​

init_pos_emb_height: int

source

Height of the initial position embedding.

init_pos_emb_time​

init_pos_emb_time: int

source

Time of the initial position embedding.

init_pos_emb_width​

init_pos_emb_width: int

source

Width of the initial position embedding.

initialize_from_config()​

classmethod initialize_from_config(pipeline_config, hf_vision_config, huggingface_config=None)

source

Initialize VisionConfig from HuggingFace vision config.

Parameters:

  • pipeline_config (PipelineConfig) – MAX Engine pipeline configuration.
  • hf_vision_config (AutoConfig) – HuggingFace vision sub-config.
  • huggingface_config (AutoConfig | None) – Full HuggingFace model config, used to derive text_hidden_size from text_config.hidden_size when hf_vision_config does not carry the attribute directly (e.g. moonshotai/Kimi-VL-A3B-Instruct vs nvidia/Kimi-K2.5-NVFP4).

Return type:

VisionConfig

Note: dtype fields will be set to defaults and should be updated via finalize() once state_dict is available.

merge_kernel_size​

merge_kernel_size: list[int]

source

Kernel size for the merge operation.

merge_type​

merge_type: str | None = None

source

Type of the merge operation.

mm_hidden_size​

mm_hidden_size: int

source

Hidden size of the multi-modal hidden layer.

mm_projector_type​

mm_projector_type: str | None = None

source

Type of the multi-modal projector.

model_type​

model_type: str = ''

source

Type of the model.

patch_size​

patch_size: int

source

Size of the patch.

pos_emb_type​

pos_emb_type: str | None = None

source

Type of the position embedding.

projector_hidden_act​

projector_hidden_act: str | None = None

source

Activation function for the projector.

projector_ln_eps​

projector_ln_eps: float

source

Epsilon for the layer normalization.

rope_max_height​

rope_max_height: int = 512

source

Maximum grid height for RoPE frequency precomputation. Hardcoded to 512 in https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/blob/main/modeling_kimi_k25.py#L571

rope_max_width​

rope_max_width: int = 512

source

Maximum grid width for RoPE frequency precomputation. Hardcoded to 512 in https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/blob/main/modeling_kimi_k25.py#L571

rope_theta​

rope_theta: float = 10000.0

source

Base for the RoPE inverse-frequency exponent. Hardcoded to 10000 in https://huggingface.co/nvidia/Kimi-K2.5-NVFP4/blob/main/modeling_kimi_k25.py#L379

text_hidden_size​

text_hidden_size: int

source

Hidden size of the text hidden layer.

video_attn_type​

video_attn_type: str | None = None

source

Type of the video attention.

vt_hidden_size​

vt_hidden_size: int

source

Hidden size of the video hidden layer.

vt_intermediate_size​

vt_intermediate_size: int

source

Intermediate size of the video hidden layer.

vt_num_attention_heads​

vt_num_attention_heads: int

source

Number of attention heads of the video hidden layer.

vt_num_hidden_layers​

vt_num_hidden_layers: int

source

Number of hidden layers of the video hidden layer.

Models

Cookbook

Puzzles

Blog

Community

Contact