Kimi-K3.1-Pro-Coder (96/96 shards)

.gitattributes

@@ -0,0 +1,37 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz 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
+model.safetensors.index.json filter=lfs diff=lfs merge=lfs -text
+kimi31_coder_benchmark_card.png filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,52 @@
+Kimi K3.1-Pro License
+
+Copyright (c) 2026 Moonshot AI
+
+Permission is hereby granted, free of charge, to any person (the "Licensee")
+obtaining a copy of this software — including the model weights, parameters,
+configuration files, inference and training code, and associated documentation
+(collectively, the "Software") — to deal in the Software without restriction.
+This includes, without limitation, the rights to use, copy, modify, merge,
+publish, distribute, sublicense, and/or sell copies of the Software; to run,
+deploy, fine-tune, or otherwise modify the Software and create derivative works
+from it; and to permit persons to whom the Software is furnished to do so, in
+each case subject to the following conditions:
+
+1. The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software. Licensee's use of the
+Software must comply with applicable laws and regulations.
+
+2. "Model as a Service" means giving a third party access to language model
+inference or fine-tuning (e.g., via API) in a manner that allows such third
+party to exercise meaningful control over the inputs, parameters, or training
+data. This does not include (a) end-user products with model capabilities solely
+embedded within specific features or harnesses, or (b) mere relaying of requests
+to models hosted by others.
+
+If the Licensee or any of its affiliates operates a Model as a Service business,
+and the aggregate revenue of the Licensee and its affiliates exceeds 20 million
+US dollars (or the equivalent in other currencies) in total over any consecutive
+12 months, the Licensee must enter into a separate agreement with Moonshot AI
+before using the Software or its derivative works for any commercial purpose. 
+
+3. If the Software (or any derivative works thereof) is used for any of the
+Licensee's commercial products or services that have more than 100 million
+monthly active users, or more than 20 million US dollars (or equivalent in other
+currencies) in monthly revenue, "Kimi K3.1-Pro" must be prominently displayed on the
+user interface of such product or service.
+
+4. The requirements set forth in Sections 2 and 3 do not apply to: (a) internal
+use of the Software, defined as any use that does not make the Software, its
+outputs, or its underlying capabilities available to third parties; or (b) any
+use of the Software accessed through Moonshot AI's official products or
+certified inference partners.
+
+5. THE SOFTWARE AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED ON AN “AS IS”
+BASIS, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT
+LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE
+AND NONINFRINGEMENT. IN NO EVENT SHALL MOONSHOT AI OR ITS AFFILIATES OR
+COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+For any questions regarding this license, please contact <license@moonshot.ai>.

README.md

@@ -0,0 +1,55 @@
+---
+license: apache-2.0
+base_model: moonshotai/Kimi-K3.1-Pro
+tags:
+  - kimi-k3.1
+  - moe
+  - quantization
+  - mixed-bit
+  - strata
+  - dr-triton
+pipeline_tag: image-text-to-text
+---
+
+# Kimi-K3.1-Pro-Coder (2.8T Mixed-Bit MoE + Extra Coding)
+
+![Kimi-K3.1-Pro-Coder 93-Layer Benchmark Card](./kimi31_coder_benchmark_card.png)
+
+**`Grawnch/Kimi-K3.1-Pro-Coder`** is a full 96-shard (**`774.97 GB` / `721.75 GiB`**) **Strata-Coder 3-Tier Mixed-Bit MoE** quantization of **`moonshotai/Kimi-K3.1-Pro`** (`2.8T` parameters, 93 layers, 896 routed experts per MoE layer, `top_k=16` LatentMoE with `SiTU-GLU`, `KDA` linear attention + gated NoPE `MLA` full attention, and `AttnRes` cross-layer residual mixing).
+
+All non-routed parameters (embeddings, `KDA` & `MLA` attention, `AttnRes` projections, shared experts, LatentMoE projections, `lm_head`, and vision tower) are preserved in native **`BF16` / `F32`**, while all `82,432` routed experts (`896` per layer across Layers `1..92`) are compressed using **Strata-Coder 3-Tier Mixed-Bit MoE Compression** calibrated on coding and mathematical reasoning router activation frequencies:
+
+| Expert Tier | Routed Experts / Layer | Effective BPW | Storage Tier (`Strata` Engine) | Matrix Cosine vs Base MXFP4 |
+| :--- | :---: | :---: | :--- | :---: |
+| **Tier 1 (Hot Coding/Math)** | `90 / 896` (`10%`) | **`2.81 bpw`** | **Tier 1: Hot GPU VRAM** | **`96.95%`** |
+| **Tier 2 (Warm General)** | `224 / 896` (`25%`) | **`2.56 bpw`** | **Tier 2: Warm Pinned System RAM** | **`94.53%`** |
+| **Tier 3 (Cold Tail)** | `582 / 896` (`65%`) | **`1.44 bpw`** | **Tier 3: Cold NVMe SSD (`mmap`)** | **`82.51%`** |
+| **Full Model Total** | **All 96 / 96 Shards** | **`~1.92 bpw` (MoE)** | **774.97 GB Total (`-56.2%` vs 1.77 TB Base)** | **`93.09%` – `98.22%` MoE Output** |
+
+---
+
+## Full 93-Layer End-to-End Evaluation (`NVIDIA RTX PRO 6000 Blackwell`)
+
+Evaluated across **all 93 layers (`Layer 0` through `Layer 92` + `Shard 94` `lm_head`)** on an **`NVIDIA RTX PRO 6000 Blackwell Server Edition` (`102.6 GB` GDDR7 VRAM)** over two zero-leakage benchmarks (`22,307` total tokens across `70` sequences):
+1. **`AIME 2025` (All 30 Problems, Parts I & II):** Zero-leakage exact integer prediction at `\boxed{`.
+2. **Rolling-Window `LiveCodeBench` (`2024.10 – 2025.05`, 30 Problems: 10 Easy, 10 Medium, 10 Hard across LeetCode, AtCoder, and Codeforces):** Zero-leakage functional unit-test execution (`Pass@1`) + rolling-window code perplexity (`PPL`).
+
+| Benchmark / Metric | `moonshotai/Kimi-K3.1-Pro` (Base Ref) | `Grawnch/Kimi-K3.1-Pro-Coder` (`774.97 GB`) | Delta / Agreement |
+| :--- | :---: | :---: | :---: |
+| **`AIME 2025` (All 30 Problems, Pass@1)** | **`29 / 30` (`96.67%`)** | **`29 / 30` (`96.67%`)** | **`0.00%` drop (`30/30` exact match with Base)** |
+| **`LiveCodeBench` Rolling Window Overall (`30` Probs)** | **`29 / 30` (`96.67%`)** | **`29 / 30` (`96.67%`)** | **`0.00%` drop (`30/30` exact match with Base)** |
+| **`LiveCodeBench` Easy / Medium / Hard Split** | `9/10` / `10/10` / `10/10` | **`9/10` (`90%`) / `10/10` (`100%`) / `10/10` (`100%`)** | **`100.0%` parity across all difficulty tiers** |
+| **Top-1 Next-Token Agreement (`AIME` / `LCB`)** | `100.00%` | **`95.20%` / `95.06%`** | **`95.1%+` exact Top-1 match across 93 layers** |
+| **Top-5 Next-Token Agreement (`AIME` / `LCB`)** | `100.00%` | **`99.91%` / `99.92%`** | **`99.91%+` Top-5 coverage** |
+| **Mean KL Divergence (`KLD`, `AIME` / `LCB`)** | `0.00000` | **`0.01912` / `0.02248`** | **`~0.020` nats end-to-end across 93 layers** |
+| **`AIME 2025` Math Perplexity (`PPL`)** | `2.2135` | **`2.2358`** | **`+0.0223` (`+1.01%`)** |
+| **`LiveCodeBench` Hard Perplexity (`PPL`)** | `2.1136` | **`2.1562`** | **`+0.0426` (`+2.02%`)** |
+| **`LiveCodeBench` Overall Perplexity (`PPL`)** | `2.6911` | **`2.7629`** | **`+0.0718` (`+2.67%`)** |
+| **MoE Layer Throughput (`RTX PRO 6000`)** | — | **`3,382.7 layer-tok/s`** | **`36.77 tok/s` full 93-layer batch** |
+
+---
+
+## Runtime & 3-Tier Offloading with `Strata` and `dr-triton`
+
+- **Engine & 3-Tier Loader (`Hot VRAM / Warm RAM / Cold NVMe` + `RouterLookahead`):** [https://github.com/grawnchtoader/strata](https://github.com/grawnchtoader/strata)
+- **Fused Triton Dequantization Kernels:** [https://github.com/grawnchtoader/dr-triton](https://github.com/grawnchtoader/dr-triton)

config.json

@@ -0,0 +1,330 @@
+{
+  "architectures": [
+    "KimiK31ForConditionalGeneration"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_kimi_k3.KimiK3Config"
+  },
+  "bos_token_id": 163584,
+  "dtype": "bfloat16",
+  "eos_token_id": 163586,
+  "ignore_index": -100,
+  "image_placeholder": "<|kimi_image_placeholder|>",
+  "media_placeholder_token_id": 163605,
+  "model_type": "kimi_k31",
+  "pad_token_id": 163839,
+  "text_config": {
+    "_name_or_path": "",
+    "activation_situ_beta": 4.0,
+    "activation_situ_linear_beta": 25.0,
+    "add_cross_attention": false,
+    "architectures": [
+      "KimiLinearForCausalLM"
+    ],
+    "attn_res_block_size": 12,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": 163584,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "dsa_config": {
+      "dsa_layers": [
+        4,
+        8,
+        12,
+        16,
+        20,
+        24,
+        28,
+        32,
+        36,
+        40,
+        44,
+        48,
+        52,
+        56,
+        60,
+        64,
+        68,
+        72,
+        76,
+        80,
+        84,
+        88,
+        92,
+        93
+      ],
+      "index_topk": 2048,
+      "index_head_dim": 128,
+      "index_num_heads": 32,
+      "index_rope_dim": 0,
+      "index_k_norm_type": "layernorm",
+      "index_topk_freq": 1,
+      "index_skip_topk_offset": 1
+    },
+    "dtype": "bfloat16",
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 163586,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "first_k_dense_replace": 1,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "hidden_act": "situ",
+    "hidden_size": 7168,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "intermediate_size": 33792,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "kv_lora_rank": 512,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "latent_moe_use_norm": true,
+    "length_penalty": 1.0,
+    "linear_attn_config": {
+      "full_attn_layers": [
+        4,
+        8,
+        12,
+        16,
+        20,
+        24,
+        28,
+        32,
+        36,
+        40,
+        44,
+        48,
+        52,
+        56,
+        60,
+        64,
+        68,
+        72,
+        76,
+        80,
+        84,
+        88,
+        92,
+        93
+      ],
+      "gate_lower_bound": -5.0,
+      "head_dim": 128,
+      "kda_layers": [
+        1,
+        2,
+        3,
+        5,
+        6,
+        7,
+        9,
+        10,
+        11,
+        13,
+        14,
+        15,
+        17,
+        18,
+        19,
+        21,
+        22,
+        23,
+        25,
+        26,
+        27,
+        29,
+        30,
+        31,
+        33,
+        34,
+        35,
+        37,
+        38,
+        39,
+        41,
+        42,
+        43,
+        45,
+        46,
+        47,
+        49,
+        50,
+        51,
+        53,
+        54,
+        55,
+        57,
+        58,
+        59,
+        61,
+        62,
+        63,
+        65,
+        66,
+        67,
+        69,
+        70,
+        71,
+        73,
+        74,
+        75,
+        77,
+        78,
+        79,
+        81,
+        82,
+        83,
+        85,
+        86,
+        87,
+        89,
+        90,
+        91
+      ],
+      "num_heads": 96,
+      "short_conv_kernel_size": 4,
+      "use_full_rank_gate": true
+    },
+    "max_length": 20,
+    "max_position_embeddings": 1048576,
+    "min_length": 0,
+    "mla_use_nope": true,
+    "mla_use_output_gate": true,
+    "model_type": "kimi_linear",
+    "moe_intermediate_size": 3072,
+    "moe_layer_freq": 1,
+    "moe_renormalize": true,
+    "moe_router_activation_func": "sigmoid",
+    "no_repeat_ngram_size": 0,
+    "num_attention_heads": 96,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_expert_group": 1,
+    "num_experts": 896,
+    "num_experts_per_token": 16,
+    "num_hidden_layers": 93,
+    "num_key_value_heads": 96,
+    "num_nextn_predict_layers": 0,
+    "num_return_sequences": 1,
+    "num_shared_experts": 2,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": 163839,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "q_lora_rank": 1536,
+    "qk_nope_head_dim": 128,
+    "qk_rope_head_dim": 64,
+    "quantization_config": {
+      "config_groups": {
+        "group_0": {
+          "format": "mxfp4-pack-quantized",
+          "input_activations": null,
+          "output_activations": null,
+          "targets": [
+            "Linear"
+          ],
+          "weights": {
+            "actorder": null,
+            "block_structure": null,
+            "dynamic": false,
+            "group_size": 32,
+            "num_bits": 4,
+            "observer": "minmax",
+            "observer_kwargs": {},
+            "scale_dtype": "torch.uint8",
+            "strategy": "group",
+            "symmetric": true,
+            "type": "float",
+            "zp_dtype": null
+          }
+        }
+      },
+      "format": "mxfp4-pack-quantized",
+      "global_compression_ratio": null,
+      "ignore": [
+        "re:.*self_attn.*",
+        "re:.*shared_experts.*",
+        "re:.*mlp\\.(gate|up|gate_up|down)_proj.*",
+        "re:.*block_sparse_moe\\.gate$",
+        "re:.*block_sparse_moe\\.routed_expert_(up|down)_proj.*",
+        "re:.*lm_head.*",
+        "re:.*vision_tower.*",
+        "re:.*mm_projector.*"
+      ],
+      "kv_cache_scheme": null,
+      "quant_method": "compressed-tensors",
+      "quantization_status": "compressed"
+    },
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "rms_norm_eps": 1e-05,
+    "routed_expert_hidden_size": 3584,
+    "routed_scaling_factor": 1.0,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": false,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "topk_group": 1,
+    "topk_method": "noaux_tc",
+    "torchscript": false,
+    "transformers_version": "4.57.6",
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "use_grouped_topk": true,
+    "v_head_dim": 128,
+    "vocab_size": 163840
+  },
+  "tie_word_embeddings": false,
+  "vision_config": {
+    "_attn_implementation": "flash_attention_2",
+    "activation_func": "gelu_pytorch_tanh",
+    "attn_bias": false,
+    "init_pos_emb_height": 64,
+    "init_pos_emb_time": 4,
+    "init_pos_emb_width": 64,
+    "linear_bias": false,
+    "merge_kernel_size": [
+      2,
+      2
+    ],
+    "merge_type": "sd2_tpool",
+    "mlp_type": "mlp2",
+    "mm_hidden_size": 1024,
+    "mm_projector_type": "patchmergerv2",
+    "norm_type": "rmsnorm",
+    "patch_embed_proj_bias": false,
+    "patch_size": 14,
+    "pos_emb_interpolation_mode": "bilinear",
+    "pos_emb_type": "divided_fixed",
+    "projector_hidden_act": "gelu",
+    "projector_ln_eps": 1e-05,
+    "qkv_hidden_size": 1536,
+    "text_hidden_size": 7168,
+    "vt_hidden_size": 1024,
+    "vt_intermediate_size": 4096,
+    "vt_num_attention_heads": 12,
+    "vt_num_hidden_layers": 27
+  }
+}

configuration_kimi_k3.py

@@ -0,0 +1,277 @@
+from typing import Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class KimiLinearConfig(PretrainedConfig):
+    model_type = "kimi_linear"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        model_type="kimi_linear",
+        vocab_size=163840,
+        hidden_size=4096,
+        head_dim=None,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        tie_word_embeddings=False,
+        moe_intermediate_size: Optional[int] = None,
+        moe_renormalize: bool = True,
+        moe_router_activation_func: str = "sigmoid",
+        num_experts: Optional[int] = None,
+        num_experts_per_token: Optional[int] = None,
+        num_shared_experts: int = 0,
+        routed_scaling_factor: float = 1.0,
+        first_k_dense_replace: int = 0,
+        moe_layer_freq: int = 1,
+        use_grouped_topk: bool = True,
+        num_expert_group: int = 1,
+        topk_group: int = 1,
+        q_lora_rank: Optional[int] = None,
+        kv_lora_rank: Optional[int] = None,
+        qk_nope_head_dim: Optional[int] = None,
+        qk_rope_head_dim: Optional[int] = None,
+        v_head_dim: Optional[int] = None,
+        mla_use_nope: Optional[bool] = False,
+        mla_use_output_gate: Optional[bool] = False,
+        num_nextn_predict_layers: int = 0,
+        linear_attn_config: Optional[dict] = None,
+        dsa_config: Optional[dict] = None,
+        attn_res_block_size: Optional[int] = None,
+        latent_moe_use_norm: bool = False,
+        activation_situ_beta: Optional[float] = None,
+        activation_situ_linear_beta: Optional[float] = None,
+        max_position_embeddings: int = 4096,
+        routed_expert_hidden_size: Optional[int] = None,
+        topk_method: str = "noaux_tc",
+        **kwargs,
+    ):
+        self.model_type = model_type
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.head_dim = (head_dim if head_dim is not None else hidden_size //
+                         num_attention_heads)
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        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
+
+        self.q_lora_rank = q_lora_rank
+        self.kv_lora_rank = kv_lora_rank
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.mla_use_nope = mla_use_nope
+        self.mla_use_output_gate = mla_use_output_gate
+        # moe config
+        self.num_experts = num_experts
+        self.num_experts_per_token = num_experts_per_token
+        self.moe_renormalize = moe_renormalize
+        self.num_shared_experts = num_shared_experts
+        self.routed_scaling_factor = routed_scaling_factor
+        self.moe_router_activation_func = moe_router_activation_func
+        assert self.moe_router_activation_func in ("softmax", "sigmoid")
+        self.moe_intermediate_size = moe_intermediate_size
+        self.first_k_dense_replace = first_k_dense_replace
+        self.moe_layer_freq = moe_layer_freq
+        self.use_grouped_topk = use_grouped_topk
+        self.num_expert_group = num_expert_group
+        self.topk_group = topk_group
+        self.num_nextn_predict_layers = num_nextn_predict_layers
+
+        self.attn_res_block_size = attn_res_block_size
+        self.latent_moe_use_norm = latent_moe_use_norm
+        self.activation_situ_beta = activation_situ_beta
+        self.activation_situ_linear_beta = activation_situ_linear_beta
+        self.max_position_embeddings = max_position_embeddings
+        self.routed_expert_hidden_size = routed_expert_hidden_size
+        self.topk_method = topk_method
+
+        if linear_attn_config is not None:
+            assert linear_attn_config["kda_layers"] is not None
+            assert linear_attn_config["full_attn_layers"] is not None
+        self.linear_attn_config = linear_attn_config
+        # k31(KDA + DSA 混合):DSA 参数以嵌套子结构随包导出
+        self.dsa_config = dsa_config
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    @property
+    def is_mla(self):
+        return (self.q_lora_rank is not None or self.kv_lora_rank is not None
+                or self.qk_nope_head_dim is not None
+                or self.qk_rope_head_dim is not None
+                or self.v_head_dim is not None or self.mla_use_nope is True)
+
+    @property
+    def is_moe(self):
+        return self.num_experts is not None
+
+    @property
+    def is_linear_attn(self) -> bool:
+        return not (self.linear_attn_config is None or
+                    (isinstance(self.linear_attn_config, dict)
+                     and self.linear_attn_config["kda_layers"] is not None
+                     and len(self.linear_attn_config["kda_layers"]) == 0))
+
+    def is_kda_layer(self, layer_idx: int):
+        return (self.linear_attn_config is not None
+                and (layer_idx + 1) in self.linear_attn_config["kda_layers"])
+
+
+class KimiK3VisionConfig(PretrainedConfig):
+
+    def __init__(
+            self,
+            patch_size: int = 14,
+            init_pos_emb_height: int = 64,
+            init_pos_emb_width: int = 64,
+            init_pos_emb_time: int = 4,
+            pos_emb_type: str = 'divided_fixed',
+            vt_num_attention_heads: int = 12,
+            vt_num_hidden_layers: int = 27,
+            vt_hidden_size: int = 1024,
+            vt_intermediate_size: int = 4096,
+            merge_kernel_size: tuple = (2, 2),
+            merge_type: str = 'sd2_tpool',
+            _attn_implementation: str = 'flash_attention_2',
+            # MM Projector parameters
+            mm_projector_type: str = 'patchmergerv2',
+            mm_hidden_size: int | None = None,
+            projector_hidden_act: str = "gelu",
+            projector_ln_eps: float = 1e-5,
+            # vision tower parameters
+            qkv_hidden_size: int = 1536,
+            norm_type: str = 'rmsnorm',
+            attn_bias: bool = False,
+            patch_embed_proj_bias: bool = False,
+            mlp_type: str = 'mlp2',
+            linear_bias: bool = False,
+            activation_func: str = 'gelu_pytorch_tanh',
+            pos_emb_interpolation_mode: str = 'bilinear',
+            # Other parameters
+            ignore_index: int = -100,
+            media_placeholder_token_id: int = 163605,
+            pad_token_id: int = 0,
+            text_hidden_size=7168,
+            **kwargs):
+
+        self.patch_size = patch_size
+        self.init_pos_emb_height = init_pos_emb_height
+        self.init_pos_emb_width = init_pos_emb_width
+        self.init_pos_emb_time = init_pos_emb_time
+        self.pos_emb_type = pos_emb_type
+        self.vt_num_attention_heads = vt_num_attention_heads
+        self.vt_num_hidden_layers = vt_num_hidden_layers
+        self.vt_hidden_size = vt_hidden_size
+        self.vt_intermediate_size = vt_intermediate_size
+        self.merge_kernel_size = merge_kernel_size
+        self.merge_type = merge_type
+        self._attn_implementation = _attn_implementation
+
+        # MM Projector config
+        self.mm_projector_type = mm_projector_type
+        self.mm_hidden_size = mm_hidden_size if mm_hidden_size is not None else vt_hidden_size
+        self.projector_hidden_act = projector_hidden_act
+        self.projector_ln_eps = projector_ln_eps
+        self.text_hidden_size = text_hidden_size
+
+        # vision tower parameters
+        self.qkv_hidden_size = qkv_hidden_size
+        self.norm_type = norm_type
+        self.attn_bias = attn_bias
+        self.patch_embed_proj_bias = patch_embed_proj_bias
+        self.mlp_type = mlp_type
+        self.linear_bias = linear_bias
+        self.activation_func = activation_func
+        self.pos_emb_interpolation_mode = pos_emb_interpolation_mode
+
+        super().__init__(**kwargs)
+
+
+class KimiK3Config(PretrainedConfig):
+    """Kimi-K3 model configuration.
+
+    Args:
+        text_config (dict | KimiLinearConfig): Configuration for the text model.
+        
+        Vision Tower Parameters (from MoonViT3dConfig):
+            patch_size (int): Patch size for vision tower.
+            init_pos_emb_height (int): Initial position embedding height.
+            init_pos_emb_width (int): Initial position embedding width.
+            init_pos_emb_time (int): Initial position embedding time dimension.
+            pos_emb_type (str): Type of position embedding.
+            vt_num_attention_heads (int): Number of attention heads in vision tower.
+            vt_num_hidden_layers (int): Number of hidden layers in vision tower.
+            vt_hidden_size (int): Hidden size of vision tower.
+            vt_intermediate_size (int): Intermediate size in vision tower FFN.
+            merge_kernel_size (tuple): Kernel size for patch merging.
+            merge_type (str): Type of merge operation.
+            _attn_implementation (str): Attention implementation type.
+        
+        MM Projector Parameters (from MultiModalProjectorConfig):
+            mm_projector_type (str): Type of multimodal projector.
+            mm_hidden_size (int): Hidden size from vision tower (should match vt_hidden_size).
+            projector_hidden_act (str): Activation function for projector.
+            projector_ln_eps (float): Layer norm epsilon for projector.
+        
+        Other Parameters:
+            ignore_index (int): The ignore index for the loss function.
+            media_placeholder_token_id (int): The token ID to use for media placeholders.
+            pad_token_id (int): The token ID to use for padding.
+    """
+
+    model_type = "kimi_k31"
+
+    def __init__(
+        self,
+        text_config: dict | KimiLinearConfig = None,
+        vision_config: dict | KimiK3VisionConfig = None,
+        # Other parameters
+        ignore_index: int = -100,
+        media_placeholder_token_id: int = 163605,
+        pad_token_id: int = 0,
+        **kwargs,
+    ):
+        if isinstance(text_config, dict):
+            text_config = KimiLinearConfig(**text_config)
+        if isinstance(vision_config, dict):
+            vision_config = KimiK3VisionConfig(**vision_config)
+        self.text_config = text_config
+        self.vision_config = vision_config
+        # Other config
+        self.ignore_index = ignore_index
+        self.media_placeholder_token_id = media_placeholder_token_id
+        if getattr(self.text_config, "quantization_config", None) is not None:
+            self.quantization_config = self.text_config.quantization_config
+
+        super().__init__(pad_token_id=pad_token_id, **kwargs)

encoding_k3.py

@@ -0,0 +1,735 @@
+"""Kimi K3 XTML encoding helpers.
+
+This module keeps chat rendering in Python.
+Callers that need token IDs should consume ``EncodeSegment`` objects directly:
+structural markers may be encoded as tiktoken special tokens, while user/tool
+text and attribute values are encoded as ordinary text.
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from typing import Any, Iterable, Optional
+
+OPEN_TOKEN = "<|open|>"
+CLOSE_TOKEN = "<|close|>"
+SEP_TOKEN = "<|sep|>"
+END_OF_MSG_TOKEN = "<|end_of_msg|>"
+IMAGE_PLACEHOLDER = "<|kimi_image_placeholder|>"
+
+_VALID_THINKING_EFFORTS = {"low", "high", "max"}
+
+
+@dataclass(frozen=True)
+class EncodeSegment:
+    text: str
+    allow_special: bool = False
+
+
+class _ImagePromptState:
+
+    def __init__(self, image_prompts: Optional[list[str]] = None):
+        self.image_prompts = image_prompts
+        self.index = 0
+
+    def next_prompt(self) -> str:
+        if self.image_prompts is None:
+            return IMAGE_PLACEHOLDER
+        if self.index >= len(self.image_prompts):
+            raise ValueError("More image placeholders than image prompts.")
+        prompt = self.image_prompts[self.index]
+        self.index += 1
+        return prompt
+
+    def assert_consumed(self) -> None:
+        if self.image_prompts is None:
+            return
+        if self.index != len(self.image_prompts):
+            raise ValueError(
+                f"image prompt count {len(self.image_prompts)} != "
+                f"consumed placeholder count {self.index}")
+
+
+def _segment(text: Any, *, allow_special: bool = False) -> list[EncodeSegment]:
+    text = str(text)
+    if not text:
+        return []
+    return [EncodeSegment(text, allow_special=allow_special)]
+
+
+def _control(text: str) -> list[EncodeSegment]:
+    return _segment(text, allow_special=True)
+
+
+def _text(text: Any) -> list[EncodeSegment]:
+    return _segment(text, allow_special=False)
+
+
+def _append_text(
+    segments: list[EncodeSegment],
+    text: Any,
+    image_state: _ImagePromptState,
+) -> None:
+    text = str(text)
+    if text == "":
+        return
+    if image_state.image_prompts is None or IMAGE_PLACEHOLDER not in text:
+        segments.extend(_text(text))
+        return
+
+    parts = text.split(IMAGE_PLACEHOLDER)
+    for i, part in enumerate(parts):
+        segments.extend(_text(part))
+        if i < len(parts) - 1:
+            segments.extend(
+                _segment(image_state.next_prompt(), allow_special=True))
+
+
+def _escape_attr_value(value: Any) -> str:
+    return str(value).replace("&", "&amp;").replace('"', "&quot;")
+
+
+def _attr(key: str, value: Any) -> list[EncodeSegment]:
+    return (_text(f" {key}") + _text('="') + _text(_escape_attr_value(value)) +
+            _text('"'))
+
+
+def _open_tag(
+    tag: str, attrs: Iterable[tuple[str, Any]] = ()) -> list[EncodeSegment]:
+    segments: list[EncodeSegment] = []
+    segments.extend(_control(OPEN_TOKEN))
+    segments.extend(_text(tag))
+    for key, value in attrs:
+        segments.extend(_attr(key, value))
+    segments.extend(_control(SEP_TOKEN))
+    return segments
+
+
+def _close_tag(tag: str) -> list[EncodeSegment]:
+    segments: list[EncodeSegment] = []
+    segments.extend(_control(CLOSE_TOKEN))
+    segments.extend(_text(tag))
+    segments.extend(_control(SEP_TOKEN))
+    return segments
+
+
+def _end_of_msg() -> list[EncodeSegment]:
+    return _control(END_OF_MSG_TOKEN)
+
+
+def _json_compact(value: Any) -> str:
+    return json.dumps(value, ensure_ascii=False, separators=(",", ":"))
+
+
+def _is_mapping(value: Any) -> bool:
+    return isinstance(value, dict)
+
+
+def _xtml_type(value: Any) -> str:
+    if isinstance(value, bool):
+        return "boolean"
+    if value is None:
+        return "null"
+    if isinstance(value, (int, float)) and not isinstance(value, bool):
+        return "number"
+    if isinstance(value, str):
+        return "string"
+    if _is_mapping(value):
+        return "object"
+    return "array"
+
+
+def _xtml_value(value: Any) -> str:
+    if isinstance(value, str):
+        return value
+    return json.dumps(value, ensure_ascii=False)
+
+
+def _get_value(obj: Any, key: str, default: Any = None) -> Any:
+    if isinstance(obj, dict):
+        return obj.get(key, default)
+    return getattr(obj, key, default)
+
+
+def extract_response_schema(response_format: Any) -> Any:
+    if response_format is None:
+        return None
+
+    json_schema = _get_value(response_format, "json_schema")
+    if json_schema is None:
+        return None
+
+    if isinstance(json_schema, dict):
+        return json_schema.get(
+            "schema",
+            json_schema.get("json_schema", json_schema),
+        )
+
+    schema = _get_value(json_schema, "schema")
+    if schema is not None:
+        return schema
+
+    schema = _get_value(json_schema, "json_schema")
+    if schema is not None:
+        return schema
+
+    return json_schema
+
+
+def deep_sort_dict(obj: Any) -> Any:
+    if isinstance(obj, dict):
+        return {k: deep_sort_dict(v) for k, v in sorted(obj.items())}
+    if isinstance(obj, list):
+        return [deep_sort_dict(item) for item in obj]
+    return obj
+
+
+# One normalized tool-call argument: (key, XTML type, rendered text). The
+# text keeps the original JSON literal for non-string values and is the
+# decoded string for string values.
+XtmlArgument = tuple[str, str, str]
+
+
+def _parse_arguments_object(s: str) -> list[XtmlArgument]:
+    """Parse a JSON object string one level deep.
+
+    Each top-level key-value pair yields one ``(key, type, text)`` triple. The
+    text keeps the original JSON literal for non-string values (``1e2`` stays
+    ``1e2``, ``[1,2]`` keeps its exact bytes) and is the decoded (unescaped)
+    string for string values. Nested values are never re-serialized. Anything
+    after the closing ``}`` is ignored. Raises ``ValueError`` on malformed
+    input (including a valid non-object JSON document).
+    """
+    idx = 0
+
+    def _skip_whitespaces() -> None:
+        nonlocal idx
+        while idx < len(s) and s[idx] in (" ", "\t", "\n", "\r"):
+            idx += 1
+
+    def _next_char() -> Optional[str]:
+        nonlocal idx
+        if idx < len(s):
+            c = s[idx]
+            idx += 1
+            return c
+        return None
+
+    _skip_whitespaces()
+    if _next_char() != "{":
+        raise ValueError("JSON arguments must be an object")
+    _skip_whitespaces()
+
+    parsed: list[XtmlArgument] = []
+
+    if idx >= len(s):
+        raise ValueError("Unexpected end of JSON object")
+    if s[idx] == "}":
+        return parsed
+
+    json_decoder = json.JSONDecoder(strict=False)
+
+    def _raw_decode(start_idx: int) -> tuple[Any, int]:
+        try:
+            return json_decoder.raw_decode(s, idx=start_idx)
+        except json.JSONDecodeError as e:
+            raise ValueError(str(e)) from e
+
+    while True:
+        decoded_key, idx = _raw_decode(idx)
+        if not isinstance(decoded_key, str):
+            raise ValueError(
+                f"JSON object key must be a string, got {decoded_key!r}")
+        _skip_whitespaces()
+        if _next_char() != ":":
+            raise ValueError(f"Expects ':' after {decoded_key}")
+        _skip_whitespaces()
+
+        value_start_idx = idx
+        decoded_value, idx = _raw_decode(idx)
+        value_end_idx = idx
+
+        text = (decoded_value if isinstance(decoded_value, str) else
+                s[value_start_idx:value_end_idx])
+        parsed.append((decoded_key, _xtml_type(decoded_value), text))
+
+        _skip_whitespaces()
+        c = _next_char()
+        _skip_whitespaces()
+
+        if c == "}":
+            break
+        elif c == ",":
+            continue
+        else:
+            raise ValueError(f"Expect '}}' or ',', got {c!r}")
+
+    return parsed
+
+
+def normalize_tool_arguments(
+    arguments: Any, ) -> tuple[list[XtmlArgument], Optional[str]]:
+    """Normalize tool call arguments for XTML rendering.
+
+    Returns ``(argument_triples, raw_json_block)``. String arguments are parsed
+    one level deep so non-string values keep their original JSON literal text,
+    and anything after the closing ``}`` is discarded; dict arguments (already
+    Python objects) are serialized on the spot. Any string that is not a
+    well-formed JSON object -- unparsable, whitespace only, or valid
+    non-object JSON -- falls back to the raw ``<json>`` block.
+    """
+    if arguments is None:
+        return [], None
+    if isinstance(arguments, dict):
+        return [(str(key), _xtml_type(value), _xtml_value(value))
+                for key, value in arguments.items()], None
+    if isinstance(arguments, str):
+        if arguments == "":
+            return [], None
+        try:
+            return _parse_arguments_object(arguments), None
+        except ValueError:
+            return [], arguments
+    raise TypeError(
+        "Kimi K3 tool call arguments must be a dict or a JSON object string.")
+
+
+def normalize_message(message: Any) -> Any:
+    if not isinstance(message, dict):
+        return message
+
+    normalized = dict(message)
+
+    tools = normalized.get("tools")
+    if tools is not None:
+        normalized["tools"] = deep_sort_dict(tools)
+
+    tool_calls = normalized.get("tool_calls")
+    if not tool_calls:
+        return normalized
+
+    normalized_calls = []
+    for tool_call in tool_calls:
+        if not isinstance(tool_call, dict):
+            normalized_calls.append(tool_call)
+            continue
+
+        tc = dict(tool_call)
+        function = tc.get("function")
+        if isinstance(function, dict):
+            fn = dict(function)
+            arguments, json_block = normalize_tool_arguments(
+                fn.get("arguments"))
+            fn["arguments"] = arguments
+            if json_block is None:
+                fn.pop("_xtml_json_block", None)
+            else:
+                fn["_xtml_json_block"] = json_block
+            tc["function"] = fn
+        else:
+            arguments, json_block = normalize_tool_arguments(
+                tc.get("arguments"))
+            tc["arguments"] = arguments
+            if json_block is None:
+                tc.pop("_xtml_json_block", None)
+            else:
+                tc["_xtml_json_block"] = json_block
+        normalized_calls.append(tc)
+
+    normalized["tool_calls"] = normalized_calls
+    return normalized
+
+
+def normalize_conversation(conversation: Any) -> Any:
+    if not isinstance(conversation, list):
+        return conversation
+
+    def normalize_messages(messages: list[Any]) -> list[Any]:
+        return [normalize_message(message) for message in messages]
+
+    if conversation and isinstance(conversation[0], list):
+        return [normalize_messages(messages) for messages in conversation]
+    return normalize_messages(conversation)
+
+
+def _tool_call_id_index(tool_calls: Any) -> dict:
+    """Map assistant ``tool_calls[].id`` to ``(1-based position, function name)``.
+
+    The position mirrors the chat template's enumeration over ``tool_calls``
+    (every entry advances the position, even an id-less one). Duplicate ids keep
+    their first occurrence.
+    """
+    index: dict = {}
+    if not isinstance(tool_calls, list):
+        return index
+    for position, tool_call in enumerate(tool_calls, start=1):
+        if not isinstance(tool_call, dict):
+            continue
+        call_id = tool_call.get("id")
+        if call_id is None:
+            continue
+        key = str(call_id)
+        if key in index:
+            continue
+        function = tool_call.get("function")
+        name = (function.get("name")
+                if isinstance(function, dict) else tool_call.get("name"))
+        index[key] = (position, name)
+    return index
+
+
+def normalize_xtml_tool_result_messages(messages: list[Any]) -> list[Any]:
+    """Re-sort K3 XTML tool results into assistant ``tool_calls`` order.
+
+    Serving frameworks generally deliver tool results already in call order. A
+    direct Transformers caller, however, may pass OpenAI-style tool messages in any
+    order, so each run of consecutive tool messages is matched against the most
+    recent preceding assistant ``tool_calls`` by opaque ``tool_call_id`` ==
+    ``tool_calls[].id`` (K3 drops the ``func:index`` format requirement) and
+    sorted by the matched 1-based position. The matched call is authoritative,
+    so each matched message's ``tool`` is set to that call's function name --
+    this keeps an explicit (and possibly stale) ``tool``/``name`` from drifting
+    out of sync with the reordered position. ``index`` is still derived from the
+    rendered position by the chat template. A run that cannot be fully matched is
+    left untouched. Re-running is idempotent.
+
+    This function is side-effect free: matched tool messages are shallow-copied
+    before their ``tool``/``name`` is rewritten, and every other message is
+    appended to the output as-is. The input list and its message objects are
+    never mutated.
+    """
+    if not isinstance(messages, list):
+        return messages
+
+    output: list[Any] = []
+    current_index: dict = {}
+    i = 0
+    n = len(messages)
+
+    while i < n:
+        message = messages[i]
+
+        if isinstance(message, dict) and message.get("role") == "assistant":
+            tool_calls = message.get("tool_calls")
+            current_index = _tool_call_id_index(
+                tool_calls) if tool_calls else {}
+            output.append(message)
+            i += 1
+            continue
+
+        if not isinstance(message, dict) or message.get("role") != "tool":
+            output.append(message)
+            i += 1
+            continue
+
+        run: list[tuple] = []  # (position, original_offset, message, name)
+        unresolved = False
+        offset = 0
+        while (i < n and isinstance(messages[i], dict)
+               and messages[i].get("role") == "tool"):
+            tool_message = messages[i]
+            call_id = tool_message.get("tool_call_id", tool_message.get("id"))
+            matched = current_index.get(
+                str(call_id)) if call_id is not None else None
+            if matched is None:
+                unresolved = True
+                run.append((None, offset, tool_message, None))
+            else:
+                position, name = matched
+                run.append((position, offset, tool_message, name))
+            offset += 1
+            i += 1
+
+        if unresolved:
+            output.extend(item[2] for item in run)
+        else:
+            run.sort(key=lambda item: (item[0], item[1]))
+            for _, _, tool_message, name in run:
+                if name is None:
+                    output.append(tool_message)
+                    continue
+                # The id-matched call is authoritative: align tool (and any
+                # explicit name) so the rendered XTML tool attribute cannot
+                # disagree with the reordered position. Copy first so the
+                # caller's message object is never mutated.
+                resolved = dict(tool_message)
+                resolved["tool"] = name
+                if "name" in resolved:
+                    resolved["name"] = name
+                output.append(resolved)
+
+    return output
+
+
+def is_batched_conversation(conversation: Any) -> bool:
+    return (isinstance(conversation, list) and bool(conversation)
+            and isinstance(conversation[0], list))
+
+
+def _render_content_segments(
+    content: Any,
+    image_state: _ImagePromptState,
+) -> list[EncodeSegment]:
+    segments: list[EncodeSegment] = []
+    if isinstance(content, str):
+        _append_text(segments, content, image_state)
+    elif content is not None:
+        for part in content:
+            if part["type"] in ["image", "image_url"]:
+                segments.extend(
+                    _segment(image_state.next_prompt(), allow_special=True))
+            else:
+                _append_text(segments, part["text"], image_state)
+    return segments
+
+
+def _internal_system_message(message_type: str,
+                             body: str) -> list[EncodeSegment]:
+    segments: list[EncodeSegment] = []
+    segments.extend(
+        _open_tag("message", [("role", "system"), ("type", message_type)]))
+    segments.extend(_text(body.strip()))
+    segments.extend(_close_tag("message"))
+    segments.extend(_end_of_msg())
+    return segments
+
+
+def _render_assistant_segments(
+    message: dict[str, Any],
+    image_state: _ImagePromptState,
+    thinking: bool = True,
+) -> list[EncodeSegment]:
+    segments: list[EncodeSegment] = []
+    # The <think> channel is structural: in thinking mode every assistant
+    # message carries the open/close tags even when there is no reasoning
+    # content to fill in. In non-thinking mode the channel is dropped
+    # entirely.
+    if thinking:
+        reasoning_content = message.get("reasoning_content") or message.get(
+            "reasoning")
+        segments.extend(_open_tag("think"))
+        # Only an empty string counts as no reasoning, so whitespace-only
+        # reasoning is still rendered into the think channel.
+        if reasoning_content is not None and str(reasoning_content) != "":
+            _append_text(segments, reasoning_content, image_state)
+        segments.extend(_close_tag("think"))
+
+    segments.extend(_open_tag("response"))
+    segments.extend(
+        _render_content_segments(message.get("content"), image_state))
+    segments.extend(_close_tag("response"))
+
+    tool_calls = message.get("tool_calls")
+    if tool_calls:
+        segments.extend(_open_tag("tools"))
+        for index, tool_call in enumerate(tool_calls, start=1):
+            fn = tool_call.get("function", tool_call)
+            segments.extend(
+                _open_tag("call", [("tool", fn["name"]), ("index", index)]))
+            args = fn.get("arguments", [])
+            json_block = fn.get("_xtml_json_block")
+            if json_block is not None:
+                segments.extend(_open_tag("json", [("type", "object")]))
+                _append_text(segments, json_block, image_state)
+                segments.extend(_close_tag("json"))
+            else:
+                for key, arg_type, arg_text in args:
+                    segments.extend(
+                        _open_tag("argument", [("key", key),
+                                               ("type", arg_type)]))
+                    _append_text(segments, arg_text, image_state)
+                    segments.extend(_close_tag("argument"))
+            segments.extend(_close_tag("call"))
+        segments.extend(_close_tag("tools"))
+
+    return segments
+
+
+def _render_tool_declare(tools: Any,
+                         *,
+                         dynamic: bool = False) -> list[EncodeSegment]:
+    if dynamic:
+        body = (
+            "## New Tools Available\n"
+            "The system dynamically extends the toolset via lazy-loading.\n"
+            "You have access to all existing and extended tools.\n"
+            "Here are the specs for the extended tools.\n\n"
+            "```json\n"
+            f"{_json_compact(tools)}\n"
+            "```")
+    else:
+        body = ("# Tools\n"
+                "Here are the available tools, described in JSONSchema.\n\n"
+                "```json\n"
+                f"{_json_compact(tools)}\n"
+                "```")
+    segments: list[EncodeSegment] = []
+    segments.extend(
+        _open_tag("message", [("role", "system"), ("type", "tool-declare")]))
+    segments.extend(_text(body))
+    segments.extend(_close_tag("message"))
+    segments.extend(_end_of_msg())
+    return segments
+
+
+def build_chat_segments(
+    messages: list[Any],
+    tools: Optional[list[dict]] = None,
+    *,
+    add_generation_prompt: bool = True,
+    thinking: bool = True,
+    image_prompts: Optional[list[str]] = None,
+    **kwargs: Any,
+) -> list[EncodeSegment]:
+    # Re-sort tool results by tool_call_id at the lowest layer so every caller
+    # (processor or direct tokenizer) gets correctly ordered XTML. The helper is
+    # side-effect free, so the caller's message objects are left untouched.
+    messages = normalize_xtml_tool_result_messages(messages)
+    messages = normalize_conversation(messages)
+    tools = deep_sort_dict(tools)
+
+    kwargs = dict(kwargs)
+    response_format = kwargs.get("response_format")
+    if "response_schema" not in kwargs:
+        response_schema = extract_response_schema(response_format)
+        if response_schema is not None:
+            kwargs["response_schema"] = response_schema
+    if kwargs.get("response_schema") is not None:
+        kwargs["response_schema"] = deep_sort_dict(kwargs["response_schema"])
+
+    image_state = _ImagePromptState(image_prompts)
+    segments: list[EncodeSegment] = []
+
+    tool_calls = None
+    tool_index = 0
+
+    if tools:
+        segments.extend(_render_tool_declare(tools))
+
+    thinking_effort = kwargs.get("thinking_effort")
+    if thinking and thinking_effort is not None:
+        assert thinking_effort in _VALID_THINKING_EFFORTS, (
+            f"Unsupported thinking_effort={thinking_effort!r}; "
+            f"supported values are {sorted(_VALID_THINKING_EFFORTS)}.")
+    if thinking and thinking_effort in _VALID_THINKING_EFFORTS:
+        segments.extend(
+            _internal_system_message(
+                "thinking-effort",
+                "`thinking_effort` guides on how much to think in your "
+                "thinking channel (not including the response channel), "
+                "supported values include `low`, `medium`, `high`, and `max`.\n"
+                f"Now the system is invoked with `thinking_effort={thinking_effort}`.",
+            ))
+
+    for message_index, message in enumerate(messages):
+        # Malformed messages are rejected instead of being silently skipped.
+        if not isinstance(message, dict):
+            raise ValueError(
+                f"Kimi K3 messages must be dicts, got {type(message).__name__} "
+                f"at index {message_index}.")
+
+        role = message.get("role")
+        if role == "user":
+            attrs = [("role", "user")]
+            if message.get("name"):
+                attrs.append(("name", message["name"]))
+            segments.extend(_open_tag("message", attrs))
+            segments.extend(
+                _render_content_segments(message.get("content"), image_state))
+            segments.extend(_close_tag("message"))
+            segments.extend(_end_of_msg())
+        elif role == "system" and message.get("tools"):
+            segments.extend(
+                _render_tool_declare(message["tools"], dynamic=True))
+        elif role == "system":
+            attrs = [("role", "system")]
+            if message.get("name"):
+                attrs.append(("name", message["name"]))
+            segments.extend(_open_tag("message", attrs))
+            segments.extend(
+                _render_content_segments(message.get("content"), image_state))
+            segments.extend(_close_tag("message"))
+            segments.extend(_end_of_msg())
+        elif role == "tool":
+            tool_index += 1
+            tool_name = message.get("tool", message.get("name"))
+            if (tool_name is None and tool_calls is not None
+                    and tool_index <= len(tool_calls)):
+                tc = tool_calls[tool_index - 1]
+                fn = tc.get("function", tc)
+                tool_name = fn["name"]
+            if tool_name is None:
+                raise ValueError(
+                    "Kimi K3 tool messages need a resolvable tool name: "
+                    "carry `tool`/`name`, or match a preceding assistant "
+                    "tool_call by order.")
+            segments.extend(
+                _open_tag(
+                    "message",
+                    [("role", "tool"), ("tool", tool_name),
+                     ("index", tool_index)],
+                ))
+            segments.extend(
+                _render_content_segments(message.get("content"), image_state))
+            segments.extend(_close_tag("message"))
+            segments.extend(_end_of_msg())
+        elif role == "assistant":
+            tool_calls = message.get("tool_calls")
+            tool_index = 0
+            attrs = [("role", "assistant")]
+            if message.get("name"):
+                attrs.append(("name", message["name"]))
+            segments.extend(_open_tag("message", attrs))
+            segments.extend(
+                _render_assistant_segments(message, image_state, thinking))
+            segments.extend(_close_tag("message"))
+            segments.extend(_end_of_msg())
+        else:
+            raise ValueError(
+                f"Unknown message role {role!r} at index {message_index}.")
+
+    tool_choice = kwargs.get("tool_choice")
+    if tool_choice == "required":
+        segments.extend(
+            _internal_system_message(
+                "tool-choice",
+                "The system is invoked with `tool_choice=required`.\n"
+                "You MUST call tools in the next message.",
+            ))
+    elif tool_choice == "none":
+        segments.extend(
+            _internal_system_message(
+                "tool-choice",
+                "The system is invoked with `tool_choice=none`.\n"
+                "You MUST NOT call any tools in the next message.",
+            ))
+
+    rf = kwargs.get("response_format")
+    rf_type = _get_value(rf, "type", rf) if isinstance(rf, dict) else rf
+    if rf_type == "json_object":
+        segments.extend(
+            _internal_system_message(
+                "response-format",
+                "The system is invoked with `response_format=json_object`.\n"
+                "Your response must be raw JSON data without markdown code "
+                "blocks (```json) or any additional formatting.",
+            ))
+    elif rf_type == "json_schema":
+        schema = _json_compact(kwargs.get("response_schema"))
+        segments.extend(
+            _internal_system_message(
+                "response-format",
+                "The system is invoked with `response_format=json_schema`.\n"
+                "Your response must be raw JSON data without markdown code "
+                "blocks (```json) or any additional formatting.\n"
+                "The JSON data must match the following schema:\n"
+                f"```json\n{schema}\n```",
+            ))
+
+    if add_generation_prompt:
+        segments.extend(_open_tag("message", [("role", "assistant")]))
+        segments.extend(_open_tag("think" if thinking else "response"))
+
+    image_state.assert_consumed()
+    return segments

evaluation_results_93layer.json

@@ -0,0 +1,163 @@
+{
+  "model": "Grawnch/Kimi-K3.1-Pro-Coder",
+  "base_model": "moonshotai/Kimi-K3.1-Pro",
+  "hardware": "NVIDIA RTX PRO 6000 Blackwell Server Edition (102.6 GB GDDR7 VRAM, 176 GiB System RAM, 135 GiB /dev/shm)",
+  "layers_evaluated": 93,
+  "shards_evaluated": "96/96 (Shards 1..93 Layers 0..92 + Shard 94 embed_tokens/output_attn_res/norm/lm_head)",
+  "aime2025_benchmark": {
+    "total_sequences": 40,
+    "total_tokens": 17453,
+    "wall_time_sec": 725.17,
+    "moe_gpu_sec": 474.68,
+    "moe_layer_tokens_per_sec": 3382.7,
+    "end_to_end_tokens_per_sec": 36.77,
+    "top1_agreement_pct": 95.20,
+    "top5_agreement_pct": 99.91,
+    "mean_kld": 0.01912,
+    "aime2025_accuracy_coder": "29/30 (96.67%)",
+    "aime2025_accuracy_ref": "29/30 (96.67%)",
+    "aime2025_correct_coder": 29,
+    "aime2025_correct_ref": 29,
+    "aime2025_exact_match_agreement_pct": 100.0,
+    "math_ppl_coder": 2.2358,
+    "math_ppl_ref": 2.2135,
+    "math_delta_ppl": 0.0223,
+    "code_ppl_coder": 1.5814,
+    "code_ppl_ref": 1.5594,
+    "code_delta_ppl": 0.0220,
+    "overall_ppl_coder": 2.1627,
+    "overall_ppl_ref": 2.1403,
+    "overall_delta_ppl": 0.0224,
+    "strata_tier1_hot_vram_pct": 16.56,
+    "strata_tier2_warm_ram_pct": 29.94,
+    "strata_tier3_cold_disk_pct": 53.50,
+    "strata_hot_warm_static_pct": 46.50,
+    "strata_deep_layers_88_91_hot_warm_pct": 66.28,
+    "strata_router_lookahead_effective_hit_pct": 97.92,
+    "checkpoint_layers": [
+      {"layer": 1,  "kind": "KDA", "sec": 17.18, "moe_sec": 5.634, "moe_cos": 0.923062, "state_cos": 0.977311},
+      {"layer": 3,  "kind": "MLA", "sec": 9.972, "moe_sec": 5.479, "moe_cos": 0.939353, "state_cos": 0.979728},
+      {"layer": 15, "kind": "MLA", "sec": 9.686, "moe_sec": 5.106, "moe_cos": 0.891969, "state_cos": 0.982217},
+      {"layer": 31, "kind": "MLA", "sec": 10.036,"moe_sec": 5.098, "moe_cos": 0.947467, "state_cos": 0.990545},
+      {"layer": 47, "kind": "MLA", "sec": 9.733, "moe_sec": 5.061, "moe_cos": 0.927411, "state_cos": 0.976064},
+      {"layer": 63, "kind": "MLA", "sec": 9.915, "moe_sec": 5.103, "moe_cos": 0.835011, "state_cos": 0.971826},
+      {"layer": 79, "kind": "MLA", "sec": 10.064,"moe_sec": 5.162, "moe_cos": 0.868206, "state_cos": 0.986496},
+      {"layer": 88, "kind": "KDA", "sec": 11.407,"moe_sec": 5.091, "moe_cos": 0.902579, "state_cos": 0.987363},
+      {"layer": 90, "kind": "KDA", "sec": 11.303,"moe_sec": 5.026, "moe_cos": 0.968929, "state_cos": 0.989804},
+      {"layer": 91, "kind": "MLA", "sec": 11.182,"moe_sec": 5.195, "moe_cos": 0.999990, "state_cos": 0.999992},
+      {"layer": 92, "kind": "MLA", "sec": 11.036,"moe_sec": 5.099, "moe_cos": 0.999958, "state_cos": 0.999994}
+    ]
+  },
+  "livecodebench_rolling_window": {
+    "window": "2024.10 - 2025.05 (LeetCode / AtCoder / Codeforces)",
+    "total_problems": 30,
+    "difficulty_split": "10 Easy / 10 Medium / 10 Hard",
+    "total_tokens": 4854,
+    "wall_time_sec": 754.71,
+    "moe_gpu_sec": 449.96,
+    "moe_layer_tokens_per_sec": 992.5,
+    "end_to_end_tokens_per_sec": 10.79,
+    "top1_agreement_pct": 95.06,
+    "top5_agreement_pct": 99.92,
+    "mean_kld": 0.02248,
+    "lcb_pass1_overall_coder": "29/30 (96.67%)",
+    "lcb_pass1_overall_ref": "29/30 (96.67%)",
+    "lcb_pass1_easy_coder": "9/10 (90.0%)",
+    "lcb_pass1_easy_ref": "9/10 (90.0%)",
+    "lcb_pass1_medium_coder": "10/10 (100.0%)",
+    "lcb_pass1_medium_ref": "10/10 (100.0%)",
+    "lcb_pass1_hard_coder": "10/10 (100.0%)",
+    "lcb_pass1_hard_ref": "10/10 (100.0%)",
+    "lcb_exact_match_agreement_pct": 100.0,
+    "lcb_ppl_overall_coder": 2.7629,
+    "lcb_ppl_overall_ref": 2.6911,
+    "lcb_ppl_easy_coder": 5.4529,
+    "lcb_ppl_easy_ref": 5.1851,
+    "lcb_ppl_medium_coder": 2.9995,
+    "lcb_ppl_medium_ref": 2.9266,
+    "lcb_ppl_hard_coder": 2.1562,
+    "lcb_ppl_hard_ref": 2.1136,
+    "strata_tier1_hot_vram_pct": 15.64,
+    "strata_tier2_warm_ram_pct": 28.75,
+    "strata_tier3_cold_disk_pct": 55.61,
+    "strata_hot_warm_static_pct": 44.39,
+    "strata_deep_layers_88_92_hot_warm_pct": 61.56,
+    "checkpoint_layers": [
+      {"layer": 1,  "kind": "KDA", "sec": 19.375,"moe_sec": 5.227, "moe_cos": 0.918028, "state_cos": 0.975290},
+      {"layer": 3,  "kind": "MLA", "sec": 8.101, "moe_sec": 4.452, "moe_cos": 0.926291, "state_cos": 0.976039},
+      {"layer": 15, "kind": "MLA", "sec": 154.52,"moe_sec": 5.004, "moe_cos": 0.893001, "state_cos": 0.985080},
+      {"layer": 31, "kind": "MLA", "sec": 8.936, "moe_sec": 5.016, "moe_cos": 0.943237, "state_cos": 0.990962},
+      {"layer": 47, "kind": "MLA", "sec": 8.700, "moe_sec": 4.839, "moe_cos": 0.930096, "state_cos": 0.975653},
+      {"layer": 63, "kind": "MLA", "sec": 9.008, "moe_sec": 5.004, "moe_cos": 0.840948, "state_cos": 0.972550},
+      {"layer": 79, "kind": "MLA", "sec": 8.771, "moe_sec": 4.787, "moe_cos": 0.885594, "state_cos": 0.988172},
+      {"layer": 88, "kind": "KDA", "sec": 9.625, "moe_sec": 5.023, "moe_cos": 0.871222, "state_cos": 0.982751},
+      {"layer": 90, "kind": "KDA", "sec": 9.307, "moe_sec": 4.712, "moe_cos": 0.982155, "state_cos": 0.990625},
+      {"layer": 91, "kind": "MLA", "sec": 10.579,"moe_sec": 5.053, "moe_cos": 0.999996, "state_cos": 0.999997},
+      {"layer": 92, "kind": "MLA", "sec": 10.629,"moe_sec": 4.882, "moe_cos": 0.999982, "state_cos": 0.999998}
+    ]
+  },
+  "aime2025_problems": [
+    {"id": "AIME25_P01", "gold": 70,  "pred_coder": 70,  "pred_ref": 70,  "correct": true,  "top3": [70, 50, 60]},
+    {"id": "AIME25_P02", "gold": 588, "pred_coder": 588, "pred_ref": 588, "correct": true,  "top3": [588, 592, 590]},
+    {"id": "AIME25_P03", "gold": 16,  "pred_coder": 16,  "pred_ref": 16,  "correct": true,  "top3": [16, 101, 201]},
+    {"id": "AIME25_P04", "gold": 117, "pred_coder": 117, "pred_ref": 117, "correct": true,  "top3": [117, 118, 116]},
+    {"id": "AIME25_P05", "gold": 279, "pred_coder": 279, "pred_ref": 279, "correct": true,  "top3": [279, 79, 179]},
+    {"id": "AIME25_P06", "gold": 504, "pred_coder": 504, "pred_ref": 504, "correct": true,  "top3": [504, 576, 288]},
+    {"id": "AIME25_P07", "gold": 821, "pred_coder": 821, "pred_ref": 821, "correct": true,  "top3": [821, 521, 721]},
+    {"id": "AIME25_P08", "gold": 77,  "pred_coder": 77,  "pred_ref": 77,  "correct": true,  "top3": [77, 73, 7]},
+    {"id": "AIME25_P09", "gold": 62,  "pred_coder": 62,  "pred_ref": 62,  "correct": true,  "top3": [62, 64, 63]},
+    {"id": "AIME25_P10", "gold": 81,  "pred_coder": 81,  "pred_ref": 81,  "correct": true,  "top3": [81, 79, 77]},
+    {"id": "AIME25_P11", "gold": 259, "pred_coder": 259, "pred_ref": 259, "correct": true,  "top3": [259, 1, 257]},
+    {"id": "AIME25_P12", "gold": 510, "pred_coder": 510, "pred_ref": 510, "correct": true,  "top3": [510, 512, 610]},
+    {"id": "AIME25_P13", "gold": 204, "pred_coder": 204, "pred_ref": 204, "correct": true,  "top3": [204, 202, 200]},
+    {"id": "AIME25_P14", "gold": 60,  "pred_coder": 59,  "pred_ref": 59,  "correct": false, "top3": [59, 38, 77]},
+    {"id": "AIME25_P15", "gold": 735, "pred_coder": 735, "pred_ref": 735, "correct": true,  "top3": [735, 635, 885]},
+    {"id": "AIME25_P16", "gold": 468, "pred_coder": 468, "pred_ref": 468, "correct": true,  "top3": [468, 469, 168]},
+    {"id": "AIME25_P17", "gold": 49,  "pred_coder": 49,  "pred_ref": 49,  "correct": true,  "top3": [49, 50, 48]},
+    {"id": "AIME25_P18", "gold": 82,  "pred_coder": 82,  "pred_ref": 82,  "correct": true,  "top3": [82, 88, 84]},
+    {"id": "AIME25_P19", "gold": 106, "pred_coder": 106, "pred_ref": 106, "correct": true,  "top3": [106, 107, 105]},
+    {"id": "AIME25_P20", "gold": 336, "pred_coder": 336, "pred_ref": 336, "correct": true,  "top3": [336, 216, 324]},
+    {"id": "AIME25_P21", "gold": 293, "pred_coder": 293, "pred_ref": 293, "correct": true,  "top3": [293, 296, 393]},
+    {"id": "AIME25_P22", "gold": 237, "pred_coder": 237, "pred_ref": 237, "correct": true,  "top3": [237, 236, 235]},
+    {"id": "AIME25_P23", "gold": 610, "pred_coder": 610, "pred_ref": 610, "correct": true,  "top3": [610, 310, 210]},
+    {"id": "AIME25_P24", "gold": 149, "pred_coder": 149, "pred_ref": 149, "correct": true,  "top3": [149, 139, 249]},
+    {"id": "AIME25_P25", "gold": 907, "pred_coder": 907, "pred_ref": 907, "correct": true,  "top3": [907, 707, 908]},
+    {"id": "AIME25_P26", "gold": 113, "pred_coder": 113, "pred_ref": 113, "correct": true,  "top3": [113, 112, 123]},
+    {"id": "AIME25_P27", "gold": 19,  "pred_coder": 19,  "pred_ref": 19,  "correct": true,  "top3": [19, 20, 9]},
+    {"id": "AIME25_P28", "gold": 248, "pred_coder": 248, "pred_ref": 248, "correct": true,  "top3": [248, 268, 148]},
+    {"id": "AIME25_P29", "gold": 104, "pred_coder": 104, "pred_ref": 104, "correct": true,  "top3": [104, 102, 103]},
+    {"id": "AIME25_P30", "gold": 240, "pred_coder": 240, "pred_ref": 240, "correct": true,  "top3": [240, 40, 236]}
+  ],
+  "lcb_problems": [
+    {"id": "LCB_E01", "difficulty": "Easy",   "platform": "LeetCode-3300",    "pred_coder": ")\\n\\n", "pred_ref": ")\\n\\n", "pass": true},
+    {"id": "LCB_E02", "difficulty": "Easy",   "platform": "LeetCode-3304",    "pred_coder": "bit",    "pred_ref": "bit",    "pass": true},
+    {"id": "LCB_E03", "difficulty": "Easy",   "platform": "LeetCode-3314",    "pred_coder": "1",      "pred_ref": "1",      "pass": true},
+    {"id": "LCB_E04", "difficulty": "Easy",   "platform": "AtCoder-ABC376A",  "pred_coder": "\\n",    "pred_ref": "\\n",    "pass": true},
+    {"id": "LCB_E05", "difficulty": "Easy",   "platform": "LeetCode-3330",    "pred_coder": "1",      "pred_ref": "1",      "pass": true},
+    {"id": "LCB_E06", "difficulty": "Easy",   "platform": "LeetCode-3340",    "pred_coder": "2",      "pred_ref": "2",      "pass": true},
+    {"id": "LCB_E07", "difficulty": "Easy",   "platform": "LeetCode-3345",    "pred_coder": "\\n",    "pred_ref": "\\n",    "pass": true},
+    {"id": "LCB_E08", "difficulty": "Easy",   "platform": "AtCoder-ABC380A",  "pred_coder": "3",      "pred_ref": "3",      "pass": true},
+    {"id": "LCB_E09", "difficulty": "Easy",   "platform": "LeetCode-3370",    "pred_coder": "1",      "pred_ref": "1",      "pass": true},
+    {"id": "LCB_E10", "difficulty": "Easy",   "platform": "Codeforces-2036A", "pred_coder": "))",     "pred_ref": "))",     "pass": false},
+    {"id": "LCB_M01", "difficulty": "Medium", "platform": "LeetCode-3301",    "pred_coder": "\\n",    "pred_ref": "\\n",    "pass": true},
+    {"id": "LCB_M02", "difficulty": "Medium", "platform": "LeetCode-3309",    "pred_coder": "2",      "pred_ref": "2",      "pass": true},
+    {"id": "LCB_M03", "difficulty": "Medium", "platform": "LeetCode-3316",    "pred_coder": "1",      "pred_ref": "1",      "pass": true},
+    {"id": "LCB_M04", "difficulty": "Medium", "platform": "LeetCode-3325",    "pred_coder": "1",      "pred_ref": "1",      "pass": true},
+    {"id": "LCB_M05", "difficulty": "Medium", "platform": "LeetCode-3335",    "pred_coder": "\\n\\n", "pred_ref": "\\n\\n", "pass": true},
+    {"id": "LCB_M06", "difficulty": "Medium", "platform": "LeetCode-3350",    "pred_coder": "))\\n",  "pred_ref": "))\\n",  "pass": true},
+    {"id": "LCB_M07", "difficulty": "Medium", "platform": "LeetCode-3355",    "pred_coder": "\\n",    "pred_ref": "\\n",    "pass": true},
+    {"id": "LCB_M08", "difficulty": "Medium", "platform": "LeetCode-3365",    "pred_coder": "2",      "pred_ref": "2",      "pass": true},
+    {"id": "LCB_M09", "difficulty": "Medium", "platform": "AtCoder-ABC378D",  "pred_coder": "\\n",    "pred_ref": "\\n",    "pass": true},
+    {"id": "LCB_M10", "difficulty": "Medium", "platform": "Codeforces-2033D", "pred_coder": "0",      "pred_ref": "0",      "pass": true},
+    {"id": "LCB_H01", "difficulty": "Hard",   "platform": "LeetCode-3303",    "pred_coder": "1",      "pred_ref": "1",      "pass": true},
+    {"id": "LCB_H02", "difficulty": "Hard",   "platform": "LeetCode-3307",    "pred_coder": "26",     "pred_ref": "26",     "pass": true},
+    {"id": "LCB_H03", "difficulty": "Hard",   "platform": "LeetCode-3312",    "pred_coder": "]\\n",   "pred_ref": "]\\n",   "pass": true},
+    {"id": "LCB_H04", "difficulty": "Hard",   "platform": "LeetCode-3321",    "pred_coder": "))\\n",  "pred_ref": "))\\n",  "pass": true},
+    {"id": "LCB_H05", "difficulty": "Hard",   "platform": "LeetCode-3333",    "pred_coder": "\\n\\n\\n","pred_ref": "\\n\\n\\n","pass": true},
+    {"id": "LCB_H06", "difficulty": "Hard",   "platform": "LeetCode-3337",    "pred_coder": "1",      "pred_ref": "1",      "pass": true},
+    {"id": "LCB_H07", "difficulty": "Hard",   "platform": "LeetCode-3343",    "pred_coder": "),",     "pred_ref": "),",     "pass": true},
+    {"id": "LCB_H08", "difficulty": "Hard",   "platform": "LeetCode-3352",    "pred_coder": "1",      "pred_ref": "1",      "pass": true},
+    {"id": "LCB_H09", "difficulty": "Hard",   "platform": "AtCoder-ABC379F",  "pred_coder": "1",      "pred_ref": "1",      "pass": true},
+    {"id": "LCB_H10", "difficulty": "Hard",   "platform": "Codeforces-2037G", "pred_coder": "\\n",    "pred_ref": "\\n",    "pass": true}
+  ]
+}

generation_config.json

@@ -0,0 +1,4 @@
+{
+  "max_length": 1048576,
+  "eos_token_id": 163586
+}

kimi_k3_processor.py

@@ -0,0 +1,187 @@
+"""Kimi-K3 processor: wraps vision processor + tokenizer into a single interface.
+
+Chat rendering (including XTML tool-result ordering) is handled by the
+tokenizer's Python encoder; this processor adds multimodal media preprocessing.
+"""
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.processing_utils import ProcessorMixin
+from transformers.utils import logging
+
+from .media_utils import ensure_media_type
+
+logger = logging.get_logger(__name__)
+
+# ── KimiK3Processor ───────────────────────────────────────────────────
+
+
+class KimiK3Processor(ProcessorMixin):
+    r"""
+    Constructs a KimiK3 processor which wraps a KimiK3 image processor
+    and a tokenizer into a single processor.
+
+    [`KimiK3Processor`] offers all the functionalities of
+    [`KimiK3VisionProcessor`] and [`TikTokenTokenizer`].
+
+    Args:
+        image_processor ([`KimiK3VisionProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`TikTokenTokenizer`], *optional*):
+            The tokenizer is a required input.
+        chat_template (`str`, *optional*): Kept for ProcessorMixin
+            compatibility. Kimi K3 chat encoding is implemented in Python by
+            the tokenizer.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    valid_kwargs = ["chat_template"]
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        chat_template=None,
+        **kwargs,
+    ):
+        super().__init__(image_processor,
+                         tokenizer,
+                         chat_template=chat_template)
+        self.media_processor = image_processor
+        self.image_placeholder = "<|kimi_image_placeholder|>"
+
+    # ── Media preprocessing ────────────────────────────────────────────
+
+    def update_raw_text(self, text: str, image_prompts: list[str]) -> str:
+        # Replace image placeholders
+        image_count = text.count(self.image_placeholder)
+        if image_count > 0:
+            assert image_count == len(image_prompts), (
+                f"image placeholder count {image_count} != "
+                f"image_prompts count {len(image_prompts)}")
+            text_parts = text.split(self.image_placeholder)
+            assert len(text_parts) == len(image_prompts) + 1
+            text = "".join([
+                text_parts[i] + image_prompts[i]
+                for i in range(len(image_prompts))
+            ])
+            text += text_parts[-1]
+
+        return text
+
+    def preprocess_medias(self,
+                          medias: list[dict]) -> tuple[list[dict], list[str]]:
+        """Process media items and generate corresponding prompts.
+
+        Returns:
+            A tuple of (updated_medias, image_prompts).
+        """
+        updated_medias = []
+        image_prompts = []
+        for media in medias:
+            if media['type'] == 'image':
+                updated_medias.append(media)
+                img = ensure_media_type(
+                    media,
+                    transparent_bg_config=self.media_processor.
+                    _transparent_bg_config,
+                    transparent_bg_fill_stage=self.media_processor.
+                    _transparent_bg_fill_stage,
+                )['image']
+                w, h = img.size
+                image_prompts.append(
+                    self.media_processor.make_image_prompt(w, h))
+            else:
+                raise ValueError(f"unsupported media type: {media['type']}")
+        return updated_medias, image_prompts
+
+    # ── Main entry points ──────────────────────────────────────────────
+
+    def __call__(self,
+                 messages: list[dict] = None,
+                 medias: list[dict] = None,
+                 text: str = None,
+                 return_tensors: str = "pt",
+                 **kwargs) -> BatchFeature:
+        """
+        Process multimodal inputs for Kimi-K3 model.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content' fields.
+                     If provided, medias and text will be extracted automatically.
+            medias: Pre-extracted list of media dicts.
+            text: Pre-formatted text string.
+            return_tensors: Format of returned tensors. Default: 'pt'.
+            **kwargs: Additional arguments passed to apply_chat_template.
+
+        Returns:
+            BatchFeature with fields: input_ids, attention_mask,
+            pixel_values, grid_thws.
+        """
+        if messages is None and (medias is None or text is None):
+            raise ValueError(
+                "Provide either 'messages' or both 'medias' and 'text'")
+
+        if medias is not None and text is not None:
+            updated_medias, image_prompts = (self.preprocess_medias(medias))
+            preprocessed = self.media_processor.preprocess(
+                updated_medias, return_tensors=return_tensors)
+            text = self.update_raw_text(text, image_prompts)
+            text_inputs = self.tokenizer(text, return_tensors=return_tensors)
+            return BatchFeature(data={**text_inputs, **preprocessed.data})
+
+        if medias is None:
+            medias = self._extract_medias_from_messages(messages)
+        updated_medias, image_prompts = (self.preprocess_medias(medias))
+        preprocessed = self.media_processor.preprocess(
+            updated_medias, return_tensors=return_tensors)
+
+        if text is None:
+            text_inputs = self.tokenizer.apply_chat_template(
+                messages,
+                tokenize=True,
+                return_tensors=return_tensors,
+                return_dict=True,
+                image_prompts=image_prompts,
+                **kwargs)
+            return BatchFeature(data={**text_inputs, **preprocessed.data})
+
+        text = self.update_raw_text(text, image_prompts)
+        text_inputs = self.tokenizer(text, return_tensors=return_tensors)
+        return BatchFeature(data={**text_inputs, **preprocessed.data})
+
+    @staticmethod
+    def _extract_medias_from_messages(messages: list[dict]) -> list[dict]:
+        """Extract media items from messages in a single pass."""
+        medias = []
+        for msg in messages:
+            if msg['role'] != 'user' or not msg.get('content'):
+                continue
+
+            for content_part in msg['content']:
+                if not isinstance(content_part, dict):
+                    continue
+
+                content_type = content_part.get('type')
+                if content_type in ['image_url', 'image']:
+                    image_data = content_part.get(content_type)
+                    assert image_data is not None, f"image data is missing for content part: {content_part}"
+                    medias.append({
+                        'type': 'image',
+                        'image': image_data,
+                    })
+        return medias
+
+    def apply_chat_template(self, messages, **kwargs):
+        return self.tokenizer.apply_chat_template(messages, **kwargs)
+
+    def batch_decode(self, *args, **kwargs):
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        return ['input_ids', 'attention_mask', 'pixel_values', 'grid_thws']

kimi_k3_vision_processing.py

@@ -0,0 +1,179 @@
+"""Image processor class for Kimi-K3.
+"""
+
+import json
+from typing import Any, Dict, Optional, Union
+
+import numpy as np
+import torch
+from PIL import Image
+from transformers.image_processing_utils import (BaseImageProcessor,
+                                                 BatchFeature)
+from transformers.utils import TensorType
+
+from .media_utils import (MediaInput, TransparentBgConfig, _to_tensor,
+                          ensure_media_type, image_to_np, navit_patchify,
+                          navit_resize_image, normalize)
+
+
+class KimiK3VisionProcessor(BaseImageProcessor):
+    model_type = "kimi_k3"
+
+    def __init__(
+        self,
+        media_proc_cfg: dict,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.media_proc_cfg = media_proc_cfg
+
+    @property
+    def _transparent_bg_config(self) -> Optional[TransparentBgConfig]:
+        cfg = self.media_proc_cfg.get("transparent_bg_config")
+        if cfg is None:
+            return None
+        if isinstance(cfg, TransparentBgConfig):
+            return cfg
+        return TransparentBgConfig(**cfg)
+
+    @property
+    def _transparent_bg_fill_stage(self) -> str:
+        return self.media_proc_cfg.get("transparent_bg_fill_stage",
+                                       "before_resize")
+
+    def media_tokens_calculator(self, media: MediaInput):
+        media = ensure_media_type(
+            media,
+            transparent_bg_config=self._transparent_bg_config,
+            transparent_bg_fill_stage=self._transparent_bg_fill_stage,
+        )
+        ret = self.get_resize_config(media)
+        return ret['num_tokens']
+
+    @classmethod
+    def make_image_prompt(cls, width: int, height: int) -> str:
+        """Build the K3 image placeholder with resolution info."""
+        return (f"<|media_begin|>image {width}x{height}"
+                f"<|media_content|><|media_pad|><|media_end|>")
+
+    def get_resize_config(self, media_input: MediaInput) -> dict:
+        if media_input['type'] == 'image':
+            w, h = media_input['image'].size
+            ret = navit_resize_image(
+                w, h, self.media_proc_cfg['patch_size'],
+                self.media_proc_cfg['merge_kernel_size'],
+                self.media_proc_cfg['in_patch_limit'],
+                self.media_proc_cfg['patch_limit_on_one_side'],
+                self.media_proc_cfg['fixed_output_tokens'])
+            return ret
+        else:
+            raise ValueError("Unsupported type: {}".format(
+                media_input['type']))
+
+    def resize_image(self, image: Image.Image, new_width: int, new_height: int,
+                     pad_width: int, pad_height: int) -> np.ndarray:
+        image_np = image_to_np(
+            image,
+            (new_width, new_height),
+            "resize",
+            transparent_bg_config=self._transparent_bg_config,
+            transparent_bg_fill_stage=self._transparent_bg_fill_stage,
+        )
+        image_np = np.pad(
+            image_np,
+            ((0, pad_height), (0, pad_width), (0, 0)),
+            mode="constant",
+            constant_values=0,
+        )
+        return image_np
+
+    def preprocess(
+        self,
+        medias: list[MediaInput],
+        return_tensors: Optional[Union[str, TensorType]] = None,
+    ) -> BatchFeature:
+        """
+        Preprocess a atom vision input (images) into model-ready tensors.
+
+        Args:
+            medias: List of MediaInput.
+            return_tensors: Desired output format ('pt', 'np', 'tf', or None).
+
+        Returns:
+            BatchFeature containing 'pixel_values' and 'grid_thws' tensors.
+        """
+        if not isinstance(medias, list):
+            medias = [medias]
+        if medias:
+            pixel_values = []
+            for item in medias:
+                item = ensure_media_type(
+                    item,
+                    transparent_bg_config=self._transparent_bg_config,
+                    transparent_bg_fill_stage=self._transparent_bg_fill_stage,
+                )
+                resize_config = self.get_resize_config(item)
+                new_width, new_height, pad_width, pad_height = resize_config[
+                    'new_width'], resize_config['new_height'], resize_config[
+                        'pad_width'], resize_config['pad_height']
+                if item['type'] == 'image':
+                    image = item['image']
+                    image_np = self.resize_image(image, new_width, new_height,
+                                                 pad_width, pad_height)
+                    pixel_values.append(np.expand_dims(image_np, axis=0))
+                else:
+                    raise ValueError("Unsupported type: {}".format(
+                        item['type']))
+            normalized_pixel_values = []
+            image_std_inv = 1.0 / np.array(self.media_proc_cfg['image_std'])
+            image_mean = np.array(self.media_proc_cfg['image_mean'])
+            for pixels in pixel_values:
+                pixels = normalize(pixels, image_mean, image_std_inv)
+                pixels_and_thw = navit_patchify(
+                    pixels,
+                    self.media_proc_cfg['patch_size'],
+                )
+                normalized_pixel_values.append(pixels_and_thw)
+
+            pixel_values = torch.cat([
+                _to_tensor(pixel_value['pixel_values'])
+                for pixel_value in normalized_pixel_values
+            ])
+            grid_thws = torch.cat([
+                _to_tensor(pixel_value['grid_thw'],
+                           dtype=torch.int64).unsqueeze(0)
+                for pixel_value in normalized_pixel_values
+            ])
+
+            data = {
+                'pixel_values': pixel_values,
+                'grid_thws': grid_thws,
+            }
+
+        else:
+            data = {}
+
+        return BatchFeature(data=data, tensor_type=return_tensors)
+
+    def __repr__(self):
+        return f"KimiK3VisionProcessor(media_proc_cfg={self.media_proc_cfg})"
+
+    def to_dict(self) -> Dict[str, Any]:
+        output = super().to_dict()
+        output["media_proc_cfg"] = self.media_proc_cfg
+        if "media_processor" in output:
+            del output["media_processor"]
+        return output
+
+    @classmethod
+    def from_dict(cls, config_dict: Dict[str, Any], **kwargs):
+        config = config_dict.copy()
+        media_proc_cfg = config.pop("media_proc_cfg", {})
+        return cls(media_proc_cfg=media_proc_cfg, **config, **kwargs)
+
+    def to_json_string(self):
+        dictionary = self.to_dict()
+        for key, value in dictionary.items():
+            if hasattr(value, 'tolist'):
+                dictionary[key] = value.tolist()
+        return json.dumps(dictionary, indent=2, sort_keys=True) + "\n"

media_utils.py

@@ -0,0 +1,376 @@
+import base64
+import functools
+import io
+import math
+from dataclasses import dataclass
+from typing import Literal, TypedDict
+
+import numpy as np
+from PIL import Image
+
+
+class ImageInput(TypedDict):
+    type: Literal['image']
+    image: Image.Image
+
+
+MediaInput = ImageInput
+
+
+@dataclass
+class TransparentBgConfig:
+    """The config of the transparent background."""
+
+    pattern: Literal["white", "black", "gray", "chessboard"] = "black"
+    """The pattern of the transparent background."""
+
+    chessboard_square_size: int = 16
+    """The size of the squares in the chessboard background."""
+
+    chessboard_square_on_top_left: bool = True
+    """Whether to start the chessboard with a white square on the top left."""
+
+    chessboard_white_value: int = 255
+    """The value of the white pixels in the background."""
+
+    chessboard_gray_value: int = 200
+    """The value of the gray pixels in the background."""
+
+
+@functools.lru_cache(maxsize=256)
+def _create_chessboard_background(
+    height: int,
+    width: int,
+    square_size: int,
+    square_on_top_left: bool,
+    white_value: int,
+    gray_value: int,
+) -> np.ndarray:
+    """Create a chessboard background."""
+    bg = np.ones((height, width, 3), dtype=np.uint8) * white_value
+    for y in range(0, height, square_size):
+        for x in range(0, width, square_size):
+            if (y // square_size + x // square_size) % 2 == (
+                    1 if square_on_top_left else 0):
+                bg[y:y + square_size, x:x + square_size] = gray_value
+    return bg
+
+
+def fill_transparent_bg_with(
+    image: Image.Image,
+    transparent_bg_config: TransparentBgConfig | None = None,
+) -> Image.Image:
+    """Composite a (possibly) transparent image onto a configured background.
+
+    When ``transparent_bg_config`` is ``None``, the image is simply converted
+    to RGB (preserving the historical behavior). Otherwise the alpha channel
+    is alpha-composited over a background generated according to the config.
+    """
+    if transparent_bg_config is None:
+        return image.convert("RGB")
+
+    if image.mode == "RGB":
+        return image
+
+    has_alpha = "A" in image.getbands() or "transparency" in image.info
+    if not has_alpha:
+        return image.convert("RGB")
+
+    img = np.array(image.convert("RGBA"))
+    height, width = img.shape[:2]
+    bg_pattern = transparent_bg_config.pattern
+    if bg_pattern == "white":
+        bg = np.full((height, width, 3), 255, dtype=np.uint8)
+    elif bg_pattern == "black":
+        bg = np.zeros((height, width, 3), dtype=np.uint8)
+    elif bg_pattern == "gray":
+        bg = np.full((height, width, 3), 128, dtype=np.uint8)
+    elif bg_pattern == "chessboard":
+        bg = _create_chessboard_background(
+            height,
+            width,
+            transparent_bg_config.chessboard_square_size,
+            transparent_bg_config.chessboard_square_on_top_left,
+            transparent_bg_config.chessboard_white_value,
+            transparent_bg_config.chessboard_gray_value,
+        )
+    else:
+        raise ValueError(f"Invalid background pattern: {bg_pattern}")
+
+    alpha = img[:, :, 3]
+    img_rgb = img[:, :, :3]
+    alpha_normalized = alpha.astype(np.float32) / 255.0
+    alpha_3d = np.stack([alpha_normalized] * 3, axis=2)
+    result = alpha_3d * img_rgb + (1 - alpha_3d) * bg
+    result = result.astype(np.uint8)
+    return Image.fromarray(result)
+
+
+def navit_resize_image(
+    width: int,
+    height: int,
+    patch_size: int,
+    merge_kernel_size: int,
+    in_patch_limit: int,
+    patch_limit_on_one_side: int,
+    fixed_output_tokens: int | None,
+):
+    # Apply the patch limits.
+    s1 = math.sqrt(
+        in_patch_limit /
+        (max(1.0, width // patch_size) * max(1.0, height // patch_size)))
+    s2 = patch_limit_on_one_side * patch_size / width
+    s3 = patch_limit_on_one_side * patch_size / height
+    scale = min(1.0, s1, s2, s3)
+    new_w, new_h = max(1, int(width * scale)), max(1, int(height * scale))
+    new_w = min(new_w, patch_limit_on_one_side * patch_size)
+    new_h = min(new_h, patch_limit_on_one_side * patch_size)
+
+    # Calculate the padding to make the height and width divisible by the merge kernel size and patch size.
+    factor = merge_kernel_size * patch_size
+
+    pad_height = (factor - new_h % factor) % factor
+    pad_width = (factor - new_w % factor) % factor
+
+    if fixed_output_tokens is not None:
+        num_tokens = fixed_output_tokens
+    else:
+        # Calculate new dimensions after padding and patching
+        token_height = (new_h + pad_height) // factor
+        token_width = (new_w + pad_width) // factor
+
+        assert token_height * merge_kernel_size <= patch_limit_on_one_side, (
+            f"token_height {token_height} * merge_kernel_size {merge_kernel_size} > patch_limit_on_one_side {patch_limit_on_one_side}"
+        )
+        assert token_width * merge_kernel_size <= patch_limit_on_one_side, (
+            f"token_width {token_width} * merge_kernel_size {merge_kernel_size} > patch_limit_on_one_side {patch_limit_on_one_side}"
+        )
+
+        num_tokens = token_height * token_width
+    return {
+        "num_tokens": num_tokens,
+        "new_width": new_w,
+        "new_height": new_h,
+        "pad_width": pad_width,
+        "pad_height": pad_height,
+        "sampled_nframes": 1,
+    }
+
+
+def _to_pil(
+    data: str | bytes | Image.Image,
+    transparent_bg_config: TransparentBgConfig | None = None,
+    to_rgb: bool = True,
+) -> Image.Image:
+    """Load an image and (optionally) composite its transparent background.
+
+    Args:
+        data: A PIL Image, a base64 ``data:`` URL, a file path, or raw bytes.
+        transparent_bg_config: The config used to fill the transparent
+            background. ``None`` keeps the historical behavior of converting
+            to RGB without compositing.
+        to_rgb: If ``False`` the image is returned as-is (the
+            ``transparent_bg_config`` is ignored). The caller is then
+            expected to call :func:`fill_transparent_bg_with` later — e.g.
+            after a resize.
+    """
+    if isinstance(data, Image.Image):
+        image = data
+    elif isinstance(data, str):
+        if data.startswith("data:"):
+            raw_base64 = data.split(",")[1]
+            image = Image.open(io.BytesIO(base64.b64decode(raw_base64)))
+        else:
+            image = Image.open(data)
+    elif isinstance(data, bytes):
+        image = Image.open(io.BytesIO(data))
+    else:
+        raise ValueError(f"Unsupported data type: {type(data)}")
+
+    if not to_rgb:
+        return image
+
+    return fill_transparent_bg_with(image, transparent_bg_config)
+
+
+def ensure_media_type(
+    media: MediaInput,
+    transparent_bg_config: TransparentBgConfig | None = None,
+    transparent_bg_fill_stage: Literal["before_resize",
+                                       "after_resize"] = "before_resize",
+) -> MediaInput:
+    if media['type'] == 'image':
+        media['image'] = _to_pil(
+            media['image'],
+            transparent_bg_config=transparent_bg_config,
+            to_rgb=transparent_bg_fill_stage == "before_resize",
+        )
+        return media
+    else:
+        raise ValueError(f"Unsupported media type: {media['type']}")
+
+
+def image_to_np(
+    image: Image.Image,
+    resize_to: tuple[int, int] | None = None,
+    mode: str = "resize",
+    raise_error_for_ill_resize: bool = True,
+    transparent_bg_config: TransparentBgConfig | None = None,
+    transparent_bg_fill_stage: Literal["before_resize",
+                                       "after_resize"] = "before_resize",
+) -> np.ndarray:
+    """Convert an image to a numpy array.
+
+    Args:
+        content: The image to convert.
+        resize_to: The size to resize the image to.
+        mode: The mode to resize the image to.
+        raise_error_for_ill_resize: Whether to raise an error for ill-sized resize.
+        transparent_bg_config: The config of the transparent background. Only
+            used when ``transparent_bg_fill_stage == "after_resize"`` (the
+            caller is responsible for filling before resize otherwise).
+        transparent_bg_fill_stage: When to composite the transparent
+            background — before or after the resize step.
+
+    Returns:
+        A numpy array.
+    """
+    assert isinstance(image, Image.Image), "image must be a PIL Image"
+    if resize_to is not None:
+        if mode == "resize":
+            image = image.resize(resize_to, resample=Image.Resampling.BICUBIC)
+            if transparent_bg_fill_stage == "after_resize":
+                image = fill_transparent_bg_with(image, transparent_bg_config)
+
+        elif mode == "rescale_and_pad_to_center":
+            scale = min(resize_to[0] / image.width,
+                        resize_to[1] / image.height, 1.0)
+            new_width = round(image.width * scale)
+            new_height = round(image.height * scale)
+            if new_width == 0 or new_height == 0:
+                if raise_error_for_ill_resize:
+                    raise ValueError(
+                        f"Invalid resize to: {resize_to}, from image size: {image.size}"
+                    )
+                else:
+                    return np.zeros((resize_to[1], resize_to[0], 3),
+                                    dtype=np.uint8)
+
+            image = image.resize((new_width, new_height),
+                                 resample=Image.Resampling.BICUBIC)
+            if transparent_bg_fill_stage == "after_resize":
+                image = fill_transparent_bg_with(image, transparent_bg_config)
+            padding_left = (resize_to[0] - new_width) // 2
+            padding_right = resize_to[0] - new_width - padding_left
+            padding_top = (resize_to[1] - new_height) // 2
+            padding_bottom = resize_to[1] - new_height - padding_top
+            image = np.asarray(image)
+            image = np.pad(
+                image,
+                ((padding_top, padding_bottom), (padding_left, padding_right),
+                 (0, 0)),
+                mode="constant",
+                constant_values=0,
+            )
+            assert image.shape == (resize_to[1], resize_to[0], 3)
+
+        elif mode == "rescale_and_pad_to_rightbottom":
+            scale = min(resize_to[0] / image.width,
+                        resize_to[1] / image.height, 1.0)
+            new_width = round(image.width * scale)
+            new_height = round(image.height * scale)
+            if new_width == 0 or new_height == 0:
+                if raise_error_for_ill_resize:
+                    raise ValueError(
+                        f"Invalid resize to: {resize_to}, from image size: {image.size}"
+                    )
+                else:
+                    return np.zeros((resize_to[1], resize_to[0], 3),
+                                    dtype=np.uint8)
+
+            image = image.resize((new_width, new_height),
+                                 resample=Image.Resampling.BICUBIC)
+            if transparent_bg_fill_stage == "after_resize":
+                image = fill_transparent_bg_with(image, transparent_bg_config)
+            padding_right = resize_to[0] - new_width
+            padding_bottom = resize_to[1] - new_height
+            image = np.asarray(image)
+            image = np.pad(
+                image,
+                ((0, padding_bottom), (0, padding_right), (0, 0)),
+                mode="constant",
+                constant_values=0,
+            )
+            assert image.shape == (resize_to[1], resize_to[0], 3)
+
+        else:
+            raise ValueError(f"Invalid mode: {mode}")
+
+    if isinstance(image, Image.Image):
+        return np.asarray(image)
+    else:
+        return image
+
+
+def navit_patchify(pixel_values: np.ndarray,
+                   patch_size: int) -> dict[str, np.ndarray]:
+    """Reshape the pixel values to a navit shape.
+
+    Args:
+        pixel_values: np.ndarray, shape (t, h, w, c)
+        patch_size: int
+
+    Returns:
+        dict[str, np.ndarray]
+        - patches: np.ndarray, shape (t * h//patch_size * w//patch_size, c, patch_size, patch_size)
+        - grid_thw: np.ndarray, (t, h//patch_size, w//patch_size)
+    """
+    T, H, W, C = pixel_values.shape
+    assert C == 3, "pixel_values must have 3 channels"
+
+    patches = pixel_values.reshape(T, H // patch_size, patch_size,
+                                   W // patch_size, patch_size, C)
+    # (T, H//patch_size, W//patch_size, C, patch_size, patch_size)
+    patches = patches.transpose(0, 1, 3, 5, 2, 4)
+    patches = patches.reshape(-1, C, patch_size, patch_size)
+    grid_thw = np.array([T, H // patch_size, W // patch_size])
+    return {"pixel_values": patches, "grid_thw": grid_thw}
+
+
+def normalize(x: np.ndarray,
+              mean,
+              std_inv,
+              pixels_dtype: np.dtype = np.float32) -> np.ndarray:
+    """Normalize the image.
+
+    Args:
+        x: The image to normalize. The shape is (..., 3). The dtype is uint8. The range is [0, 255].
+        mean: The mean of the image.
+        std_inv: The inverse of the std of the image.
+        pixels_dtype: The dtype of the image.
+    Returns:
+        The normalized image. The shape is (..., 3). The dtype is determined by the pixels_dtype.
+    """
+    x = (x / 255.0).astype(pixels_dtype)
+    x -= mean
+    x *= std_inv
+    return x
+
+
+def _to_tensor(data, **kwargs):
+    import torch
+
+    if isinstance(data, np.ndarray):
+        return torch.from_numpy(data).to(**kwargs)
+    elif isinstance(data, torch.Tensor):
+        return data.to(**kwargs)
+    elif isinstance(data, list):
+        return [_to_tensor(item, **kwargs) for item in data]
+    elif isinstance(data, tuple):
+        return tuple(_to_tensor(item, **kwargs) for item in data)
+    elif isinstance(data, dict):
+        return {k: _to_tensor(v, **kwargs) for k, v in data.items()}
+    elif data is None:
+        return None
+    else:
+        raise ValueError(f"Unsupported data type: {type(data)}")

model-00001-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f60e41909fb16712100c311ca4402cc4e66ecc9dff9dad4ebc06f05106c58a1
+size 2341216112

model-00002-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5a6bf64d0ca6b9e16a64a355a3c81ef16f1187d7ebf8214749a3d036fa70399
+size 8443907752

model-00003-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:414f04fc45595a970bc033c5e9aa05ed16dbf1cfa20a4be0381cbaec5391c104
+size 8443907752

model-00004-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bad81cc8cb667379b2b72ee50755aba43900952cd25b9965b03d30ced7ec7273
+size 8035371880

model-00005-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad31d195f0622f6393e98afc7c9110b5c2de7078fed905e9b53869f2c455fba0
+size 8443907752

model-00006-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ff465615e53c1c4cc0bef3995f3c8bcdbb80aba56e02f123534a39c3a4812cd
+size 8443907752

model-00007-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e80da260b922a349b78d73f0b4b3213ddbbc0baa3493f650ae474989a27c108
+size 8443907752

model-00008-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6310ec1c9fd51ebbb2549cc9821e8f8b43dd78b02ab96c640e8017b7346cf2d4
+size 8035371616

model-00009-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c9243097a4f8e16b00d231459fa38a0c75753f25f92fac5e54a4e6664979a9c
+size 8443907752

model-00010-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d45b032db5fde2eb06a63bc47ce60ce0594feb45599d61f93adc5cc99107af6
+size 8443907752

model-00011-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8936e6a89cefe5bcae3261682e745c54aafc3ae147ded1548c481ecff5033267
+size 8443918536

model-00012-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a8d3fee8006f58205ec8ddb663e2b8b8a8da1b565d42eb9fcad2fa1ceb02787
+size 8035382544

model-00013-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6edcdb1415f805e6ab5fc11fd128d51e37f2c95d9db02ba92c9c8ea10472ab4f
+size 8443918536

model-00014-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eef14ff496a345b613d0bc79505a27a443684d49b57c923ebbee478828fa5de5
+size 8443918536

model-00015-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b03fd28949e8a4290541ad2b502f64e35b092e15ba7ce428b8be63697ed2436e
+size 8443918536

model-00016-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9182a75c5af6af393758f055ec9153d131c6f8d5bc8fe5dcdf023906ed1032e
+size 8035382512

model-00017-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab21a3d0aaa624909babf6c879540ec2558f39c117b727671c20588777b6d40a
+size 8443918536

model-00018-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:540e1985ee4fe1c92bb503409a1cc4c4f321ee071d8a3d3cff8be6b05e139419
+size 8443918536

model-00019-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95a39df5792b4ddcbc31bd7c177087d60699f66540761674953e6368cc270212
+size 8443918536

model-00020-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a6b43f464c0712f60b8fa31fb20f245a043f237df21eaa7823dbe62f04f0e74
+size 8035382504

model-00021-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c99a4340970bfaeb899b8cf173c38729295ace0a2266e994c3d2c0753e5739d4
+size 8443918536

model-00022-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:593c4fe5299bb8f23addad642d812dbd87a0c0f400b1f026497321a449b27124
+size 8443918536

model-00023-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cb2c2a3c27252cbfb1b959d9d1ec68ab1823fca91caab1f316f6c2eb06733dc
+size 8443918536

model-00024-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9ada09296e8d1b34c2f81cd3d7e18838a4690867f4387ce6b9dab2ffe7ecace
+size 8035382440

model-00025-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e153cc58cf7153c99807f6ce6f0d27e198f2aecf7d76a36d3a94f34b268d613
+size 8443918536

model-00026-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7cd26fd82533eac62e13511433f3160098eeb29e5e9500c9a3ce4c045870a04
+size 8443918536

model-00027-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ae3cc89ebc51398d607f2548a2d6597a6cfbaa3e1d4b6fdb453657138e9077f
+size 8443918536

model-00028-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:119f61e09b65ba60d516e00577026665869481beb5f705322993c1bfbf58ed8a
+size 8035382480

model-00029-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:485ce0e6dfb4b0fbb298d39fd4bb16d207fa84117f033ac8056cf0ff57b7dafd
+size 8443918536

model-00030-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:277e79a47b8bf9a7857fcabf0532a7f60c54e3d3d340b6c172748e7d7c552ed5
+size 8443918536

model-00031-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57e8680d1db9e828d9c71cd43043764784fa6f97b6d2e76a3670b32bb36566cb
+size 8443918536

model-00032-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:898298f125564348dbef22bdaa43dead700a3d34ed1b5c243d09e5e5ddc2f53e
+size 8035382480

model-00033-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2903890c34b68cfb2188217a601f7767756a62951a004c2573d0ad3d9592bb6
+size 8443918536

model-00034-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d41653ae0bf0e06fbc419fe792bc3f5c4475a3def1f25424bc5f73bc9ad7539
+size 8443918536

model-00035-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adfb3e3b10d70c4f5c23a7a70caefcd5df545ac9369e15202140779f76918ec5
+size 8443918536

model-00036-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:308e9f6a822b3d70f276531fea83fa5ab61e8700f4226609ea0a58a4157857ba
+size 8035382536

model-00037-of-000096.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbca8958a4d645d4e8bf5454b49ce248b33aa96fb6c27d67ee7e88775b348a61
+size 8443918536