Merge branch 'master' into compilade/refactor-kv-cache

Also begin reverting some implicit state rollback code.

gguf-py/gguf/constants.py

@@ -94,9 +94,12 @@ class Keys:
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
         ATTN_LOGIT_SOFTCAPPING            = "{arch}.attn_logit_softcapping"
         FINAL_LOGIT_SOFTCAPPING           = "{arch}.final_logit_softcapping"
+        SWIN_NORM                         = "{arch}.swin_norm"
         RESCALE_EVERY_N_LAYERS            = "{arch}.rescale_every_n_layers"
         TIME_MIX_EXTRA_DIM                = "{arch}.time_mix_extra_dim"
         TIME_DECAY_EXTRA_DIM              = "{arch}.time_decay_extra_dim"
+        RESIDUAL_SCALE                    = "{arch}.residual_scale"
+        EMBEDDING_SCALE                   = "{arch}.embedding_scale"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"
@@ -112,6 +115,7 @@ class Keys:
         KV_LORA_RANK      = "{arch}.attention.kv_lora_rank"
         REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count"
         SLIDING_WINDOW    = "{arch}.attention.sliding_window"
+        SCALE             = "{arch}.attention.scale"
 
     class Rope:
         DIMENSION_COUNT         = "{arch}.rope.dimension_count"
@@ -148,6 +152,8 @@ class Keys:
         MERGES               = "tokenizer.ggml.merges"
         BOS_ID               = "tokenizer.ggml.bos_token_id"
         EOS_ID               = "tokenizer.ggml.eos_token_id"
+        EOT_ID               = "tokenizer.ggml.eot_token_id"
+        EOM_ID               = "tokenizer.ggml.eom_token_id"
         UNK_ID               = "tokenizer.ggml.unknown_token_id"
         SEP_ID               = "tokenizer.ggml.seperator_token_id"
         PAD_ID               = "tokenizer.ggml.padding_token_id"
@@ -164,11 +170,16 @@ class Keys:
         CHAT_TEMPLATE_N      = "tokenizer.chat_template.{name}"
         CHAT_TEMPLATES       = "tokenizer.chat_templates"
         # FIM/Infill special tokens constants
+        FIM_PRE_ID           = "tokenizer.ggml.fim_pre_token_id"
+        FIM_SUF_ID           = "tokenizer.ggml.fim_suf_token_id"
+        FIM_MID_ID           = "tokenizer.ggml.fim_mid_token_id"
+        FIM_PAD_ID           = "tokenizer.ggml.fim_pad_token_id"
+        FIM_REP_ID           = "tokenizer.ggml.fim_rep_token_id"
+        FIM_SEP_ID           = "tokenizer.ggml.fim_sep_token_id"
+        # deprecated:
         PREFIX_ID            = "tokenizer.ggml.prefix_token_id"
         SUFFIX_ID            = "tokenizer.ggml.suffix_token_id"
         MIDDLE_ID            = "tokenizer.ggml.middle_token_id"
-        EOT_ID               = "tokenizer.ggml.eot_token_id"
-        EOM_ID               = "tokenizer.ggml.eom_token_id"
 
     class Adapter:
         TYPE       = "adapter.type"
@@ -210,6 +221,7 @@ class MODEL_ARCH(IntEnum):
     ORION        = auto()
     INTERNLM2    = auto()
     MINICPM      = auto()
+    MINICPM3     = auto()
     GEMMA        = auto()
     GEMMA2       = auto()
     STARCODER2   = auto()
@@ -220,6 +232,7 @@ class MODEL_ARCH(IntEnum):
     COMMAND_R    = auto()
     DBRX         = auto()
     OLMO         = auto()
+    OLMOE        = auto()
     OPENELM      = auto()
     ARCTIC       = auto()
     DEEPSEEK2    = auto()
@@ -230,6 +243,9 @@ class MODEL_ARCH(IntEnum):
     JAIS         = auto()
     NEMOTRON     = auto()
     EXAONE       = auto()
+    GRANITE      = auto()
+    GRANITE_MOE  = auto()
+    CHAMELEON    = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -340,6 +356,8 @@ class MODEL_TENSOR(IntEnum):
     ENC_FFN_DOWN         = auto()
     ENC_FFN_UP           = auto()
     ENC_OUTPUT_NORM      = auto()
+    CLS                  = auto() # classifier
+    CLS_OUT              = auto() # classifier output projection
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -368,6 +386,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.ORION:          "orion",
     MODEL_ARCH.INTERNLM2:      "internlm2",
     MODEL_ARCH.MINICPM:        "minicpm",
+    MODEL_ARCH.MINICPM3:       "minicpm3",
     MODEL_ARCH.GEMMA:          "gemma",
     MODEL_ARCH.GEMMA2:         "gemma2",
     MODEL_ARCH.STARCODER2:     "starcoder2",
@@ -378,6 +397,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.COMMAND_R:      "command-r",
     MODEL_ARCH.DBRX:           "dbrx",
     MODEL_ARCH.OLMO:           "olmo",
+    MODEL_ARCH.OLMOE:          "olmoe",
     MODEL_ARCH.OPENELM:        "openelm",
     MODEL_ARCH.ARCTIC:         "arctic",
     MODEL_ARCH.DEEPSEEK2:      "deepseek2",
@@ -388,6 +408,9 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.JAIS:           "jais",
     MODEL_ARCH.NEMOTRON:       "nemotron",
     MODEL_ARCH.EXAONE:         "exaone",
+    MODEL_ARCH.GRANITE:        "granite",
+    MODEL_ARCH.GRANITE_MOE:    "granitemoe",
+    MODEL_ARCH.CHAMELEON:      "chameleon",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -498,6 +521,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.ENC_FFN_DOWN:              "enc.blk.{bid}.ffn_down",
     MODEL_TENSOR.ENC_FFN_UP:                "enc.blk.{bid}.ffn_up",
     MODEL_TENSOR.ENC_OUTPUT_NORM:           "enc.output_norm",
+    MODEL_TENSOR.CLS:                       "cls",
+    MODEL_TENSOR.CLS_OUT:                   "cls.output",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -607,6 +632,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.LAYER_OUT_NORM,
+        MODEL_TENSOR.CLS,
+        MODEL_TENSOR.CLS_OUT,
     ],
     MODEL_ARCH.NOMIC_BERT: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -638,6 +665,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_GATE,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.LAYER_OUT_NORM,
+        MODEL_TENSOR.CLS,
     ],
     MODEL_ARCH.MPT: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -801,6 +829,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FACTORS_LONG,
+        MODEL_TENSOR.ROPE_FACTORS_SHORT,
         MODEL_TENSOR.ATTN_NORM,
         MODEL_TENSOR.ATTN_QKV,
         MODEL_TENSOR.ATTN_Q,
@@ -875,6 +905,25 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
+    MODEL_ARCH.MINICPM3: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FACTORS_LONG,
+        MODEL_TENSOR.ROPE_FACTORS_SHORT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q_A,
+        MODEL_TENSOR.ATTN_Q_B,
+        MODEL_TENSOR.ATTN_KV_A_MQA,
+        MODEL_TENSOR.ATTN_KV_B,
+        MODEL_TENSOR.ATTN_Q_A_NORM,
+        MODEL_TENSOR.ATTN_KV_A_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     MODEL_ARCH.GEMMA: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -1044,6 +1093,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.OLMOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+    ],
     MODEL_ARCH.OPENELM: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -1222,6 +1288,51 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.GRANITE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.GRANITE_MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
+    MODEL_ARCH.CHAMELEON: [
+        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_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 
@@ -1511,6 +1622,8 @@ KEY_TOKENIZER_SCORES     = Keys.Tokenizer.SCORES
 KEY_TOKENIZER_MERGES     = Keys.Tokenizer.MERGES
 KEY_TOKENIZER_BOS_ID     = Keys.Tokenizer.BOS_ID
 KEY_TOKENIZER_EOS_ID     = Keys.Tokenizer.EOS_ID
+KEY_TOKENIZER_EOT_ID     = Keys.Tokenizer.EOT_ID
+KEY_TOKENIZER_EOM_ID     = Keys.Tokenizer.EOM_ID
 KEY_TOKENIZER_UNK_ID     = Keys.Tokenizer.UNK_ID
 KEY_TOKENIZER_SEP_ID     = Keys.Tokenizer.SEP_ID
 KEY_TOKENIZER_PAD_ID     = Keys.Tokenizer.PAD_ID
@@ -1518,8 +1631,15 @@ KEY_TOKENIZER_CLS_ID     = Keys.Tokenizer.CLS_ID
 KEY_TOKENIZER_MASK_ID    = Keys.Tokenizer.MASK_ID
 KEY_TOKENIZER_HF_JSON    = Keys.Tokenizer.HF_JSON
 KEY_TOKENIZER_RWKV       = Keys.Tokenizer.RWKV
-KEY_TOKENIZER_PRIFIX_ID  = Keys.Tokenizer.PREFIX_ID
+
+KEY_TOKENIZER_FIM_PRE_ID = Keys.Tokenizer.FIM_PRE_ID
+KEY_TOKENIZER_FIM_SUF_ID = Keys.Tokenizer.FIM_SUF_ID
+KEY_TOKENIZER_FIM_MID_ID = Keys.Tokenizer.FIM_MID_ID
+KEY_TOKENIZER_FIM_PAD_ID = Keys.Tokenizer.FIM_PAD_ID
+KEY_TOKENIZER_FIM_REP_ID = Keys.Tokenizer.FIM_REP_ID
+KEY_TOKENIZER_FIM_SEP_ID = Keys.Tokenizer.FIM_SEP_ID
+
+# deprecated
+KEY_TOKENIZER_PREFIX_ID  = Keys.Tokenizer.PREFIX_ID
 KEY_TOKENIZER_SUFFIX_ID  = Keys.Tokenizer.SUFFIX_ID
 KEY_TOKENIZER_MIDDLE_ID  = Keys.Tokenizer.MIDDLE_ID
-KEY_TOKENIZER_EOT_ID     = Keys.Tokenizer.EOT_ID
-KEY_TOKENIZER_EOM_ID     = Keys.Tokenizer.EOM_ID

gguf-py/gguf/gguf_writer.py

@@ -670,6 +670,9 @@ class GGUFWriter:
     def add_expert_weights_scale(self, value: float) -> None:
         self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
 
+    def add_swin_norm(self, value: bool) -> None:
+        self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value)
+
     def add_rescale_every_n_layers(self, count: int) -> None:
         self.add_uint32(Keys.LLM.RESCALE_EVERY_N_LAYERS.format(arch=self.arch), count)
 
@@ -679,6 +682,12 @@ class GGUFWriter:
     def add_time_decay_extra_dim(self, dim: int) -> None:
         self.add_uint32(Keys.LLM.TIME_DECAY_EXTRA_DIM.format(arch=self.arch), dim)
 
+    def add_residual_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.RESIDUAL_SCALE.format(arch=self.arch), value)
+
+    def add_embedding_scale(self, value: float) -> None:
+        self.add_float32(Keys.LLM.EMBEDDING_SCALE.format(arch=self.arch), value)
+
     def add_wkv_head_size(self, size: int) -> None:
         self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size)
 
@@ -703,6 +712,9 @@ class GGUFWriter:
     def add_sliding_window(self, value: int) -> None:
         self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value)
 
+    def add_attention_scale(self, value: float) -> None:
+        self.add_float32(Keys.Attention.SCALE.format(arch=self.arch), value)
+
     def add_pooling_type(self, value: PoolingType) -> None:
         self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)
 
@@ -831,15 +843,6 @@ class GGUFWriter:
 
         self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value)
 
-    def add_prefix_token_id(self, id: int) -> None:
-        self.add_uint32(Keys.Tokenizer.PREFIX_ID, id)
-
-    def add_suffix_token_id(self, id: int) -> None:
-        self.add_uint32(Keys.Tokenizer.SUFFIX_ID, id)
-
-    def add_middle_token_id(self, id: int) -> None:
-        self.add_uint32(Keys.Tokenizer.MIDDLE_ID, id)
-
     def add_eot_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOT_ID, id)
 

gguf-py/gguf/tensor_mapping.py

@@ -13,7 +13,7 @@ class TensorNameMap:
             "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx jais exaone
             "transformer.word_embeddings",               # falcon
             "word_embeddings",                           # bloom
-            "model.embed_tokens",                        # llama-hf nemotron
+            "model.embed_tokens",                        # llama-hf nemotron olmoe
             "tok_embeddings",                            # llama-pth
             "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
@@ -54,7 +54,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
@@ -66,7 +66,7 @@ class TensorNameMap:
         MODEL_TENSOR.OUTPUT_NORM: (
             "gpt_neox.final_layer_norm",               # gptneox
             "transformer.ln_f",                        # gpt2 gpt-j falcon jais exaone
-            "model.norm",                              # llama-hf baichuan internlm2
+            "model.norm",                              # llama-hf baichuan internlm2 olmoe
             "norm",                                    # llama-pth
             "transformer.norm_f",                      # mpt dbrx
             "ln_f",                                    # refact bloom qwen gpt2
@@ -87,6 +87,9 @@ class TensorNameMap:
             "rope.freqs",  # llama-pth
             "rotary_pos_emb.inv_freq",  # chatglm
         ),
+
+        MODEL_TENSOR.ROPE_FACTORS_LONG: (),
+        MODEL_TENSOR.ROPE_FACTORS_SHORT: (),
     }
 
     block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
@@ -98,7 +101,7 @@ class TensorNameMap:
             "transformer.h.{bid}.input_layernorm",                  # falcon7b
             "h.{bid}.input_layernorm",                              # bloom
             "transformer.h.{bid}.ln_mlp",                           # falcon40b
-            "model.layers.{bid}.input_layernorm",                   # llama-hf nemotron
+            "model.layers.{bid}.input_layernorm",                   # llama-hf nemotron olmoe
             "layers.{bid}.attention_norm",                          # llama-pth
             "language_model.encoder.layers.{bid}.input_layernorm",  # persimmon
             "model.layers.{bid}.ln1",                               # yi
@@ -142,7 +145,7 @@ class TensorNameMap:
 
         # Attention query
         MODEL_TENSOR.ATTN_Q: (
-            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron
+            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe
             "layers.{bid}.attention.wq",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.query",                  # bert
             "transformer.h.{bid}.attn.q_proj",                           # gpt-j
@@ -154,7 +157,7 @@ class TensorNameMap:
 
         # Attention key
         MODEL_TENSOR.ATTN_K: (
-            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron
+            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe
             "layers.{bid}.attention.wk",                               # llama-pth
             "encoder.layer.{bid}.attention.self.key",                  # bert
             "transformer.h.{bid}.attn.k_proj",                         # gpt-j
@@ -167,7 +170,7 @@ class TensorNameMap:
 
         # Attention value
         MODEL_TENSOR.ATTN_V: (
-            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron
+            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe
             "layers.{bid}.attention.wv",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.value",                  # bert
             "transformer.h.{bid}.attn.v_proj",                           # gpt-j
@@ -185,7 +188,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.attn.out_proj",                       # mpt
             "transformer.h.{bid}.self_attention.dense",                     # falcon
             "h.{bid}.self_attention.dense",                                 # bloom
-            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron
+            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe
             "layers.{bid}.attention.wo",                                    # llama-pth
             "encoder.layer.{bid}.attention.output.dense",                   # bert
             "transformer.h.{bid}.attn.out_proj",                            # gpt-j
@@ -229,7 +232,7 @@ class TensorNameMap:
             "transformer.h.{bid}.ln_2",                                      # gpt2 refact qwen jais exaone
             "h.{bid}.post_attention_layernorm",                              # bloom
             "transformer.blocks.{bid}.norm_2",                               # mpt
-            "model.layers.{bid}.post_attention_layernorm",                   # llama-hf nemotron
+            "model.layers.{bid}.post_attention_layernorm",                   # llama-hf nemotron olmoe
             "layers.{bid}.ffn_norm",                                         # llama-pth
             "language_model.encoder.layers.{bid}.post_attention_layernorm",  # persimmon
             "model.layers.{bid}.ln2",                                        # yi
@@ -253,12 +256,13 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_GATE_INP: (
-            "layers.{bid}.feed_forward.gate",             # mixtral
-            "model.layers.{bid}.block_sparse_moe.gate",   # mixtral
-            "model.layers.{bid}.mlp.gate",                # qwen2moe
-            "transformer.decoder_layer.{bid}.router",     # Grok
-            "transformer.blocks.{bid}.ffn.router.layer",  # dbrx
-            "model.layers.{bid}.feed_forward.router",     # jamba
+            "layers.{bid}.feed_forward.gate",                   # mixtral
+            "model.layers.{bid}.block_sparse_moe.gate",         # mixtral
+            "model.layers.{bid}.mlp.gate",                      # qwen2moe olmoe
+            "transformer.decoder_layer.{bid}.router",           # Grok
+            "transformer.blocks.{bid}.ffn.router.layer",        # dbrx
+            "model.layers.{bid}.feed_forward.router",           # jamba
+            "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
         ),
 
         MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
@@ -299,7 +303,7 @@ class TensorNameMap:
             "layers.{bid}.feed_forward.experts.w3",          # mixtral (merged)
             "transformer.decoder_layer.{bid}.moe.linear_v",  # Grok (merged)
             "transformer.blocks.{bid}.ffn.experts.mlp.v1",   # dbrx
-            "model.layers.{bid}.mlp.experts.up_proj",        # qwen2moe (merged)
+            "model.layers.{bid}.mlp.experts.up_proj",        # qwen2moe olmoe (merged)
         ),
 
         MODEL_TENSOR.FFN_UP_SHEXP: (
@@ -332,7 +336,7 @@ class TensorNameMap:
             "layers.{bid}.feed_forward.experts.w1",         # mixtral (merged)
             "transformer.decoder_layer.{bid}.moe.linear",   # Grok (merged)
             "transformer.blocks.{bid}.ffn.experts.mlp.w1",  # dbrx
-            "model.layers.{bid}.mlp.experts.gate_proj",     # qwen2moe (merged)
+            "model.layers.{bid}.mlp.experts.gate_proj",     # qwen2moe olmoe (merged)
         ),
 
         MODEL_TENSOR.FFN_GATE_SHEXP: (
@@ -370,10 +374,11 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
-            "layers.{bid}.feed_forward.experts.w2",          # mixtral (merged)
-            "transformer.decoder_layer.{bid}.moe.linear_1",  # Grok (merged)
-            "transformer.blocks.{bid}.ffn.experts.mlp.w2",   # dbrx
-            "model.layers.{bid}.mlp.experts.down_proj",      # qwen2moe (merged)
+            "layers.{bid}.feed_forward.experts.w2",              # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear_1",      # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.w2",       # dbrx
+            "model.layers.{bid}.mlp.experts.down_proj",          # qwen2moe olmoe (merged)
+            "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
         ),
 
         MODEL_TENSOR.FFN_DOWN_SHEXP: (
@@ -384,7 +389,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.q_norm",                            # cohere
+            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
             "transformer.layers.{bid}.attn.q_norm",                           # openelm
@@ -393,7 +398,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.k_norm",                            # cohere
+            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
             "transformer.layers.{bid}.attn.k_norm",                           # openelm
@@ -704,6 +709,15 @@ class TensorNameMap:
         MODEL_TENSOR.ENC_OUTPUT_NORM: (
             "encoder.final_layer_norm", # t5
         ),
+
+        MODEL_TENSOR.CLS: (
+            "classifier",       # jina
+            "classifier.dense", # roberta
+        ),
+
+        MODEL_TENSOR.CLS_OUT: (
+            "classifier.out_proj", # roberta
+        ),
     }
 
     # architecture-specific block mappings

gguf-py/gguf/vocab.py

@@ -122,8 +122,30 @@ class SpecialVocab:
                 tokenizer = json.load(f)
             if self.load_merges:
                 merges = tokenizer.get('model', {}).get('merges')
-                if isinstance(merges, list) and merges and isinstance(merges[0], str):
-                    self.merges = merges
+                if isinstance(merges, list) and merges:
+                    if isinstance(merges[0], str):
+                        self.merges = merges
+                    elif isinstance(merges[0], list) and len(merges[0]) == 2 and isinstance(merges[0][0], str):
+                        # New format since transformers 4.45 to support spaces in merges
+                        # ref: https://github.com/ggerganov/llama.cpp/issues/9692
+                        # TODO: internally store as the new format instead of converting to old
+                        if any(' ' in s for pair in merges for s in pair):
+                            logger.warning(f'Spaces in merges detected, encoding as {chr(ord(" ") + 256)!r}')
+                        self.merges = [
+                            ' '.join(
+                                [
+                                    # ensure the spaces are properly encoded
+                                    ''.join(
+                                        chr(ord(c) + 256) if c == ' ' else c
+                                        for c in part
+                                    )
+                                    for part in pair
+                                ]
+                            )
+                            for pair in merges
+                        ]
+                    else:
+                        raise ValueError("Unknown tokenizer merges format")
             added_tokens = tokenizer.get('added_tokens', {})
         else:
             added_tokens = {}