[Migration] Migrate GGUF quantization support to plugin

.github/dependabot.yml

@@ -21,7 +21,6 @@ updates:
       - dependency-name: "torchvision"
       - dependency-name: "xformers"
       - dependency-name: "lm-format-enforcer"
-      - dependency-name: "gguf"
       - dependency-name: "compressed-tensors"
       - dependency-name: "ray[cgraph]" # Ray Compiled Graph
       - dependency-name: "lm-eval"

.pre-commit-config.yaml

@@ -21,7 +21,7 @@ repos:
   rev: v21.1.2
   hooks:
   - id: clang-format
-    exclude: 'csrc/(moe/topk_softmax_kernels.cu|libtorch_stable/quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
+    exclude: 'csrc/moe/topk_softmax_kernels.cu|vllm/third_party/.*'
     types_or: [c++, cuda]
     args: [--style=file, --verbose]
 - repo: https://github.com/DavidAnson/markdownlint-cli2

CMakeLists.txt

@@ -631,7 +631,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
     "csrc/libtorch_stable/quantization/w8a8/int8/per_token_group_quant.cu"
     "csrc/libtorch_stable/permute_cols.cu"
     "csrc/libtorch_stable/quantization/gptq/q_gemm.cu"
-    "csrc/libtorch_stable/quantization/gguf/gguf_kernel.cu"
     "csrc/libtorch_stable/pos_encoding_kernels.cu"
     "csrc/libtorch_stable/fused_qknorm_rope_kernel.cu"
     "csrc/libtorch_stable/layernorm_kernels.cu"

csrc/libtorch_stable/ops.h

@@ -397,35 +397,6 @@ torch::stable::Tensor gptq_gemm(torch::stable::Tensor a,
 void gptq_shuffle(torch::stable::Tensor q_weight, torch::stable::Tensor q_perm,
                   int64_t bit);
 
-// GGML kernels (shared CUDA/ROCm)
-torch::stable::Tensor ggml_dequantize(
-    torch::stable::Tensor W, int64_t type, int64_t m, int64_t n,
-    std::optional<torch::headeronly::ScalarType> const& dtype);
-
-torch::stable::Tensor ggml_mul_mat_vec_a8(torch::stable::Tensor W,
-                                          torch::stable::Tensor X, int64_t type,
-                                          int64_t row);
-
-torch::stable::Tensor ggml_mul_mat_a8(torch::stable::Tensor W,
-                                      torch::stable::Tensor X, int64_t type,
-                                      int64_t row);
-
-torch::stable::Tensor ggml_moe_a8(torch::stable::Tensor X,
-                                  torch::stable::Tensor W,
-                                  torch::stable::Tensor sorted_token_ids,
-                                  torch::stable::Tensor expert_ids,
-                                  torch::stable::Tensor num_tokens_post_padded,
-                                  int64_t type, int64_t row, int64_t top_k,
-                                  int64_t tokens);
-
-torch::stable::Tensor ggml_moe_a8_vec(torch::stable::Tensor X,
-                                      torch::stable::Tensor W,
-                                      torch::stable::Tensor topk_ids,
-                                      int64_t top_k, int64_t type, int64_t row,
-                                      int64_t tokens);
-
-int64_t ggml_moe_get_block_size(int64_t type);
-
 void paged_attention_v1(
     torch::stable::Tensor& out, torch::stable::Tensor& query,
     torch::stable::Tensor& key_cache, torch::stable::Tensor& value_cache,