Chunk 12.0

In this chunk, the assistant executed a decisive pivot from the failed NVFP4 quantization path toward deploying GLM-5 via GGUF UD-Q4_K_XL on vLLM. After freeing disk space and beginning the 431 GB GGUF download, the assistant discovered a critical blocker: vLLM's GGUF support depends on `transformers` for architecture metadata parsing, and neither `transformers` (v5.2.0) nor the installed `gguf-py` (0.17.1) included the `glm-dsa` architecture used by GLM-5. Multiple GitHub issues confirmed that DeepSeek/GLM GGUF models consistently fail with `ValueError: GGUF model with architecture deepseek2 is not supported yet`. The user explicitly rejected all alternative paths (reverting to sglang, llama.cpp, or FP8) and directed the assistant to **add GGUF support to vLLM** (option E). The assistant launched parallel deep-dive research into three codebases: `transformers` GGUF config mapping, vLLM's `GGUFModelLoader`, and the GLM-5 GGUF tensor structure via the latest `gguf-py` from llama.cpp HEAD. Key findings included: (1) the blocker is solely in `transformers` — vLLM already has manual weight mappings for DeepSeek architectures; (2) the `gguf-py` library from llama.cpp already defines `LLM_ARCH_GLM_DSA` with a complete tensor name map; (3) the GGUF file splits `kv_b_proj` into separate `attn_k_b` and `attn_v_b` tensors that must be reassembled; (4) expert weights use fused `gate_up_proj` format and `e_score_correction_bias` needs manual mapping. The assistant installed vLLM nightly (`0.16.0rc2.dev313`), upgraded `transformers` to `5.3.0.dev0` from git HEAD, installed `gguf-py` from llama.cpp source (adding `glm-dsa` architecture support), and began writing a comprehensive patch for vLLM's `gguf_loader.py` to handle `glm_moe_dsa` model_type with expert weight sideloading, KV split reassembly, and indexer/nextn tensor mapping. The GGUF download was started in the background, and the assistant is now positioned to apply and test the patch once the download completes.

The Great GGUF Pivot: Engineering vLLM Support for GLM-5 from Scratch 3996 words

Message Articles

Subagent Sessions