llama: accept either glm5next indexer kpool key name

The two competing GLM-5.3-Flash drafts write the same value under
different GGUF keys: ggml-org#27752 uses attention.indexer.block_size, ggml-org#27754
uses attention.indexer.kpool. Semantics and the
indexer_top_k % kpool == 0 constraint are identical, so read the
block_size key first and fall back to kpool instead of aborting at
"GLM5NEXT requires index_kpool" on a GGUF from the other converter.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: 2 people

gguf-py/gguf/constants.py

@@ -222,6 +222,7 @@ class Indexer:
             KEY_LENGTH = "{arch}.attention.indexer.key_length"
             TOP_K      = "{arch}.attention.indexer.top_k"
             BLOCK_SIZE   = "{arch}.attention.indexer.block_size"    # MSA
+            KPOOL        = "{arch}.attention.indexer.kpool"         # glm5next alias (PR #27754)
             LOCAL_BLOCKS = "{arch}.attention.indexer.local_blocks"  # MSA
             TYPES      = "{arch}.attention.indexer.types"
 

src/llama-arch.cpp

@@ -282,6 +282,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,           "%s.attention.indexer.key_length"           },
     { LLM_KV_ATTENTION_INDEXER_TOP_K,                "%s.attention.indexer.top_k"                },
     { LLM_KV_ATTENTION_INDEXER_BLOCK_SIZE,           "%s.attention.indexer.block_size"           },
+    { LLM_KV_ATTENTION_INDEXER_KPOOL,                "%s.attention.indexer.kpool"                },
     { LLM_KV_ATTENTION_INDEXER_LOCAL_BLOCKS,         "%s.attention.indexer.local_blocks"         },
     { LLM_KV_ATTENTION_INDEXER_TYPES,                "%s.attention.indexer.types"                },
     { LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT,           "%s.attention.output_group_count"           },

src/llama-arch.h

@@ -287,6 +287,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,
     LLM_KV_ATTENTION_INDEXER_TOP_K,
     LLM_KV_ATTENTION_INDEXER_BLOCK_SIZE,
+    LLM_KV_ATTENTION_INDEXER_KPOOL,
     LLM_KV_ATTENTION_INDEXER_LOCAL_BLOCKS,
     LLM_KV_ATTENTION_INDEXER_TYPES,
     LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT,

src/models/glm5next.cpp

@@ -73,8 +73,15 @@ void llama_model_glm5next::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,  hparams.indexer_n_head,     false);
     ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,  hparams.indexer_head_size,  false);
     ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K,       hparams.indexer_top_k,      false);
-    // index_kpool: tokens per compressed key, always_select_tail is implied (see build_dsa_top_k)
+    // index_kpool: tokens per compressed key, always_select_tail is implied (see build_dsa_top_k).
+    // Converters disagree on the key name for this value: upstream PR #27752 writes
+    // `attention.indexer.block_size`, the competing PR #27754 writes `attention.indexer.kpool`.
+    // The semantics and the `indexer_top_k % kpool == 0` constraint are identical, so accept
+    // either rather than aborting on a GGUF produced by the other converter.
     ml.get_key(LLM_KV_ATTENTION_INDEXER_BLOCK_SIZE,  hparams.indexer_block_size, false);
+    if (hparams.indexer_block_size == 0) {
+        ml.get_key(LLM_KV_ATTENTION_INDEXER_KPOOL,   hparams.indexer_block_size, false);
+    }
 
     if (hparams.indexer_head_size > 0) {
         GGML_ASSERT(hparams.indexer_n_head > 0);