Support for dots.llm1 models (ikawrakow#573)

* Add llama.cpp changes for dots1 support

* Add python changes for dots1 support

* Fix to make it convert

* Remove V reshaping, remove BOS by default for dots1 and fix warmup to handle models without BOS

* Minor fix

* Remove commented lines

Author: saood06

convert_hf_to_gguf.py

@@ -3864,6 +3864,34 @@ def prepare_tensors(self):
         self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias)
 
 
+@Model.register("Dots1ForCausalLM")
+class Dots1Model(Qwen2MoeModel):
+    model_arch = gguf.MODEL_ARCH.DOTS1
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["num_experts"] = self.hparams["n_routed_experts"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
+        self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
+
+        if self.hparams["scoring_func"] == "sigmoid":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        else:
+            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+        if "shared_experts" in name:
+            return [(self.map_tensor_name(name), data_torch)]
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @Model.register("ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(Model):
     model_arch = gguf.MODEL_ARCH.CHATGLM

gguf-py/gguf/constants.py

@@ -226,6 +226,7 @@ class MODEL_ARCH(IntEnum):
     T5           = auto()
     T5ENCODER    = auto()
     JAIS         = auto()
+    DOTS1        = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -362,6 +363,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.T5:             "t5",
     MODEL_ARCH.T5ENCODER:      "t5encoder",
     MODEL_ARCH.JAIS:           "jais",
+    MODEL_ARCH.DOTS1:          "dots1",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -1164,6 +1166,30 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_GATE,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.DOTS1: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+    ],
     # TODO
 }
 

gguf-py/gguf/tensor_mapping.py

@@ -257,7 +257,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_EXP_PROBS_B: (
-            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
+            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
         ),
 
         # Feed-forward up