From 9b1a1b91faef673b40836c8b42f937f8719efbe7 Mon Sep 17 00:00:00 2001 From: Saba Fallah <10401143+sfallah@users.noreply.github.com> Date: Wed, 11 Feb 2026 14:10:11 +0100 Subject: [PATCH] minor --- convert_hf_to_gguf.py | 5 ++--- gguf-py/gguf/constants.py | 1 - 2 files changed, 2 insertions(+), 4 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 17959dde2f..9e659b98b5 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -7758,15 +7758,14 @@ class DeepseekV2Model(TextModel): # Default: if no MoE, all layers are dense; if MoE, none are dense first_k_dense_replace = hparams["num_hidden_layers"] if not has_moe else 0 self.gguf_writer.add_leading_dense_block_count(first_k_dense_replace) - kv_lora_rank = hparams["kv_lora_rank"] if hparams.get("kv_lora_rank") is not None else 512 + kv_lora_rank = hparams.get("kv_lora_rank", 512) self.gguf_writer.add_vocab_size(hparams["vocab_size"]) if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None: self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"]) - if "kv_lora_rank" in hparams and hparams["kv_lora_rank"] is not None: - self.gguf_writer.add_kv_lora_rank(kv_lora_rank) # note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA if not is_ocr: + self.gguf_writer.add_kv_lora_rank(kv_lora_rank) self.gguf_writer.add_key_length(kv_lora_rank + hparams["qk_rope_head_dim"]) self.gguf_writer.add_value_length(kv_lora_rank) self.gguf_writer.add_key_length_mla(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"]) diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 3cc4967611..c3c1d40ff7 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -300,7 +300,6 @@ class Keys: IMAGE_MEAN = "clip.vision.image_mean" IMAGE_STD = "clip.vision.image_std" SPATIAL_MERGE_SIZE = "clip.vision.spatial_merge_size" - WINDOW_SIZE = "clip.vision.window_size" USE_GELU = "clip.use_gelu" USE_SILU = "clip.use_silu" N_WA_PATTERN = "clip.vision.n_wa_pattern" # used by qwen2.5vl