convert : BailingMoE : fix qkv split when head_dim is 0

[CISC]

Missed second head_dim usage in #12678.

Cleaner assignment as well.

Edit: The Ling-lite-base model is still broken though until PR#2 is merged.

@bartowski1182 @nicoboss

[nicoboss]

Thanks a lot for fixing this so quickly! I tested it and can confirm that this fixes any issues I experianced with Ling-lite-base GGUF convearsion. However unfortinately the resulting GGUF doesn't seam to load in llama.cpp:

root@AI:/apool/llama.cpp/build/bin# ./llama-cli -m /mradermacher/tmp/quant/Ling-lite-base.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 2 CUDA devices:
  Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
  Device 1: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
build: 5016 (fb8c6eb4) with cc (Debian 12.2.0-14) 12.2.0 for x86_64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 4090) - 23241 MiB free
llama_model_load_from_file_impl: using device CUDA1 (NVIDIA GeForce RTX 4090) - 23260 MiB free
llama_model_loader: loaded meta data with 38 key-value pairs and 367 tensors from /mradermacher/tmp/quant/Ling-lite-base.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = bailingmoe
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Ling Lite Base
llama_model_loader: - kv   3:                         general.size_label str              = 64x1.5B
llama_model_loader: - kv   4:                            general.license str              = mit
llama_model_loader: - kv   5:                               general.tags arr[str,1]       = ["text-generation"]
llama_model_loader: - kv   6:                     bailingmoe.block_count u32              = 28
llama_model_loader: - kv   7:                  bailingmoe.context_length u32              = 16384
llama_model_loader: - kv   8:                bailingmoe.embedding_length u32              = 2048
llama_model_loader: - kv   9:             bailingmoe.feed_forward_length u32              = 5632
llama_model_loader: - kv  10:            bailingmoe.attention.head_count u32              = 16
llama_model_loader: - kv  11:         bailingmoe.attention.head_count_kv u32              = 4
llama_model_loader: - kv  12:                  bailingmoe.rope.freq_base f32              = 600000.000000
llama_model_loader: - kv  13: bailingmoe.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  14:               bailingmoe.expert_used_count u32              = 6
llama_model_loader: - kv  15:            bailingmoe.attention.key_length u32              = 0
llama_model_loader: - kv  16:          bailingmoe.attention.value_length u32              = 0
llama_model_loader: - kv  17:                          general.file_type u32              = 1
llama_model_loader: - kv  18:            bailingmoe.rope.dimension_count u32              = 128
llama_model_loader: - kv  19:               bailingmoe.rope.scaling.type str              = none
llama_model_loader: - kv  20:       bailingmoe.leading_dense_block_count u32              = 0
llama_model_loader: - kv  21:                      bailingmoe.vocab_size u32              = 126464
llama_model_loader: - kv  22:      bailingmoe.expert_feed_forward_length u32              = 1408
llama_model_loader: - kv  23:            bailingmoe.expert_weights_scale f32              = 1.000000
llama_model_loader: - kv  24:                    bailingmoe.expert_count u32              = 64
llama_model_loader: - kv  25:             bailingmoe.expert_shared_count u32              = 2
llama_model_loader: - kv  26:             bailingmoe.expert_weights_norm bool             = true
llama_model_loader: - kv  27:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  28:                         tokenizer.ggml.pre str              = bailingmoe
llama_model_loader: - kv  29:                      tokenizer.ggml.tokens arr[str,126464]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  30:                  tokenizer.ggml.token_type arr[i32,126464]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  31:                      tokenizer.ggml.merges arr[str,125824]  = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "h e...
llama_model_loader: - kv  32:                tokenizer.ggml.bos_token_id u32              = 126080
llama_model_loader: - kv  33:                tokenizer.ggml.eos_token_id u32              = 126081
llama_model_loader: - kv  34:            tokenizer.ggml.padding_token_id u32              = 126081
llama_model_loader: - kv  35:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  36:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  37:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   85 tensors
llama_model_loader: - type  f16:  282 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = F16
print_info: file size   = 31.30 GiB (16.00 BPW) 
load: special tokens cache size = 262
load: token to piece cache size = 0.8056 MB
print_info: arch             = bailingmoe
print_info: vocab_only       = 0
print_info: n_ctx_train      = 16384
print_info: n_embd           = 2048
print_info: n_layer          = 28
print_info: n_head           = 16
print_info: n_head_kv        = 4
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 0
print_info: n_embd_head_v    = 0
print_info: n_gqa            = 4
print_info: n_embd_k_gqa     = 0
print_info: n_embd_v_gqa     = 0
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-05
print_info: f_clamp_kqv      = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale    = 0.0e+00
print_info: f_attn_scale     = 0.0e+00
print_info: n_ff             = 5632
print_info: n_expert         = 64
print_info: n_expert_used    = 6
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 0
print_info: rope scaling     = none
print_info: freq_base_train  = 600000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 16384
print_info: rope_finetuned   = unknown
print_info: ssm_d_conv       = 0
print_info: ssm_d_inner      = 0
print_info: ssm_d_state      = 0
print_info: ssm_dt_rank      = 0
print_info: ssm_dt_b_c_rms   = 0
print_info: model type       = 16B
print_info: model params     = 16.80 B
print_info: general.name     = Ling Lite Base
print_info: n_layer_dense_lead   = 0
print_info: n_ff_exp             = 1408
print_info: n_expert_shared      = 2
print_info: expert_weights_scale = 1.0
print_info: expert_weights_norm  = 1
print_info: vocab type       = BPE
print_info: n_vocab          = 126464
print_info: n_merges         = 125824
print_info: BOS token        = 126080 '<|startoftext|>'
print_info: EOS token        = 126081 '<|endoftext|>'
print_info: EOT token        = 126081 '<|endoftext|>'
print_info: PAD token        = 126081 '<|endoftext|>'
print_info: LF token         = 198 'Ċ'
print_info: EOG token        = 126081 '<|endoftext|>'
print_info: max token length = 154
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 0 repeating layers to GPU
load_tensors: offloaded 0/29 layers to GPU
load_tensors:   CPU_Mapped model buffer size = 32054.45 MiB
...........................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: freq_base     = 600000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (16384) -- the full capacity of the model will not be utilized
llama_context:        CPU  output buffer size =     0.48 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 28, can_shift = 1
init: failed to allocate buffer for kv cache
llama_init_from_model: failed to initialize the context: failed to initialize self-attention cache
common_init_from_params: failed to create context with model '/mradermacher/tmp/quant/Ling-lite-base.gguf'
main: error: unable to load model

[CISC]

Thanks a lot for fixing this so quickly! I tested it and can confirm that this fixes any issues I experianced with Ling-lite-base GGUF convearsion. However unfortinately the resulting GGUF doesn't seam to load in llama.cpp:

Sigh, indeed, it's because the base class does this:

llama.cpp/convert_hf_to_gguf.py

Lines 257 to 259 in a6f32f0

	if (head_dim := self.hparams.get("head_dim")) is not None:
	self.gguf_writer.add_key_length(head_dim)
	self.gguf_writer.add_value_length(head_dim)

I'll see if we can work around that.

[CISC]

I'll see if we can work around that.

Actually, you know what, setting head_dim to 0 is broken, I'll submit a fix to the model instead.

See PR#2

[CISC]

@ngxson I still think it's worth it to merge this as it's slightly nicer, even though the model itself is what needs fixing.

[ngxson]

I don't have time to test this model, but seems good to me

[ngxson]

Yeah sometimes we have issue where models exported by tranformers has some keys set to None, will discuss with @huggingface team to see if it can be removed in next version

[ngxson]

@CISC I checked with transformers team, they said that the None value is actually set by a custom code outside of the library.

More importantly, I almost forgot that we actually have Model.find_hparams in convert_hf_to_gguf.py that is perfect for handling such case. If you have time, can you do a pass to change places that currently using .get to find_hparams?

[CISC]

That method is useful if you have multiple candidates for a value, but I don't see how it applies here?

The issue is not None, but that they set an actual 0.

[ngxson]

Hmm ok I misunderstood the function. But now I think it would be nice if find_hparam can handle this case, maybe add an arg default_value to it?

I'm thinking about this because it was also the case for gemma 3, is was a bit messy because some params are either missing or being null in the config.json, could be possible that many models in the future will have this same behavior.