[Mini-VLLM] Cut non-OpenAI surfaces, dead kernels, Mistral stack
- Remove Anthropic, Cohere, SageMaker, decisions, structured-decisions, rl, render-launcher and batch entrypoints; keep scale-out token routes - Drop the Swagger static bundle and the /docs, /redoc routes - Delete device-keyed tuning configs for fused_moe, quant utils, mamba select-state and helion; no sm86 entry ever existed for any of them - Remove ten unused rope variants plus the fbgemm_fp8 and fp_quant quantizers and their config, metrics and loader plumbing - Cut the CPU device build: csrc/cpu, cpu_extension.cmake, cpu requirements and the setup.py cpu target path - Delete 29 environment variables nothing reads and 14 chat-template fallbacks for models this fork no longer registers - Remove nine dead CUDA ops (deepseek_v4, kimi_k3, minimax, muse_spark, dsv3 gemm) with their kernels, bindings, wrappers and CMake entries - Delete the connector test suites that targeted removed connectors - Cut the Mistral tokenizer and parser stack, its tests, and the mistral_common and ijson dependency pins - Restore the Qwen4Exp registry rows and nvidia/mtp.py that an earlier cut deleted while keeping the model tree Co-authored-by: Hinata <hinata@idihore.id> Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
@@ -318,12 +318,6 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
|
||||
"awq_dequantize(Tensor _kernel, Tensor _scaling_factors, "
|
||||
"Tensor _zeros, SymInt split_k_iters, int thx, int thy) -> Tensor");
|
||||
|
||||
// DeepSeek V3 fused A GEMM (SM 9.0+, bf16 only, 1-16 tokens).
|
||||
// conditionally compiled so impl registration is in source file
|
||||
ops.def(
|
||||
"dsv3_fused_a_gemm(Tensor! output, Tensor mat_a, Tensor mat_b, "
|
||||
"bool enable_pdl=False) -> ()");
|
||||
|
||||
// BF16/FP32 x FP32 -> FP32 router GEMM for H=3072, E=256, M<=32 (SM90+).
|
||||
// conditionally compiled so impl registration is in source file
|
||||
ops.def("fp32_router_gemm(Tensor! output, Tensor mat_a, Tensor mat_b) -> ()");
|
||||
@@ -410,133 +404,6 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
|
||||
"bool is_neox, Tensor position_ids, "
|
||||
"int forced_token_heads_per_warp=-1) -> ()");
|
||||
|
||||
// q_head_padded is the padded Q head count of the returned tensor, or 0 to
|
||||
// do the KV insert alone and return an empty tensor. The Q knobs are
|
||||
// independent: apply_q_norm and apply_q_rope each drop that step for Q
|
||||
// alone (KV is always rotated), and is_q_interleaved reads and writes Q in
|
||||
// FlashMLA's mega-attention chunk-interleaved layout, which moves the
|
||||
// padding heads to the tail of every head-dim chunk.
|
||||
ops.def(
|
||||
"fused_deepseek_v4_kv_rope_insert("
|
||||
"Tensor kv, Tensor! k_cache, Tensor slot_mapping, Tensor position_ids, "
|
||||
"Tensor cos_sin_cache, int cache_block_size, Tensor? fp8_scale=None, "
|
||||
"bool kv_mxfp8=False) -> "
|
||||
"()");
|
||||
ops.def(
|
||||
"fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert("
|
||||
"Tensor q_in, Tensor kv, Tensor! k_cache, "
|
||||
"Tensor slot_mapping, Tensor position_ids, Tensor cos_sin_cache, "
|
||||
"int q_head_padded, float eps, int cache_block_size, "
|
||||
"bool apply_q_norm=True, bool kv_mxfp8=False, bool apply_q_rope=True, "
|
||||
"bool is_q_interleaved=False) -> Tensor");
|
||||
|
||||
// FlashInfer V4 full-cache variants: write Q in place (bf16) or to a separate
|
||||
// FP8 tensor, and KV into a contiguous 512-wide token-strided cache.
|
||||
ops.def(
|
||||
"fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert("
|
||||
"Tensor! q, Tensor kv, Tensor! k_cache, Tensor slot_mapping, "
|
||||
"Tensor position_ids, Tensor cos_sin_cache, float eps, "
|
||||
"int cache_block_size, bool apply_q_norm=True) -> ()");
|
||||
ops.def(
|
||||
"fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert("
|
||||
"Tensor q, Tensor kv, Tensor! q_fp8, Tensor! k_cache, "
|
||||
"Tensor slot_mapping, Tensor position_ids, Tensor cos_sin_cache, "
|
||||
"Tensor fp8_scale, Tensor q_fp8_scale_inv, float eps, "
|
||||
"int cache_block_size, bool apply_q_norm=True) -> ()");
|
||||
|
||||
// Kimi-K3 MLA epilogues: optional RoPE followed by concat/cache insertion.
|
||||
ops.def(
|
||||
"fused_kimi_k3_mla_key_concat_kv_cache_insert("
|
||||
"Tensor! q, Tensor k_nope, Tensor k_pe, Tensor kv_c_normed, "
|
||||
"Tensor! k_out, Tensor! k_cache, Tensor slot_mapping, "
|
||||
"int cache_block_size, Tensor? position_ids=None, "
|
||||
"Tensor? cos_sin_cache=None) -> ()");
|
||||
ops.def(
|
||||
"fused_kimi_k3_mla_key_concat_ds_mla_insert("
|
||||
"Tensor! q, Tensor k_nope, Tensor k_pe, Tensor kv_c_normed, "
|
||||
"Tensor! k_out, Tensor! k_cache, Tensor slot_mapping, "
|
||||
"int cache_block_size, Tensor? position_ids=None, "
|
||||
"Tensor? cos_sin_cache=None) -> ()");
|
||||
ops.def(
|
||||
"fused_kimi_k3_mla_kv_concat(Tensor k_nope, Tensor k_pe, Tensor! k_out) "
|
||||
"-> ()");
|
||||
ops.def(
|
||||
"fused_kimi_k3_mla_kv_concat_quant_fp8("
|
||||
"Tensor k_nope, Tensor k_pe, Tensor v, Tensor! k_fp8, Tensor! v_fp8) "
|
||||
"-> ()");
|
||||
ops.def(
|
||||
"fused_kimi_k3_mla_qkv_quant_kv_cache_fp8_insert("
|
||||
"Tensor q, Tensor k_nope, Tensor k_pe, Tensor kv_c_normed, Tensor v, "
|
||||
"Tensor! q_fp8, Tensor! k_fp8, Tensor! v_fp8, Tensor! k_cache, "
|
||||
"Tensor slot_mapping, Tensor q_scale_inv, Tensor k_scale_inv, "
|
||||
"Tensor v_scale_inv, Tensor cache_scale_inv, int cache_block_size, "
|
||||
"Tensor? position_ids=None, Tensor? cos_sin_cache=None) -> ()");
|
||||
|
||||
// Kimi-K3 MLA decode epilogue: concat mqa_q = [ql_nope | q_pe] and insert the
|
||||
// latent [kv_c_normed | k_pe] into the paged cache (bf16 / fp8 / fp8_ds_mla).
|
||||
ops.def(
|
||||
"fused_kimi_k3_mla_decode_q_concat_kv_cache_insert("
|
||||
"Tensor ql_nope, Tensor q_pe, Tensor kv_c_normed, Tensor k_pe, "
|
||||
"Tensor! mqa_q, Tensor! k_cache, Tensor slot_mapping, "
|
||||
"int cache_block_size, Tensor? position_ids=None, "
|
||||
"Tensor? cos_sin_cache=None) -> ()");
|
||||
ops.def(
|
||||
"fused_kimi_k3_mla_decode_q_concat_kv_cache_fp8_insert("
|
||||
"Tensor ql_nope, Tensor q_pe, Tensor kv_c_normed, Tensor k_pe, "
|
||||
"Tensor! mqa_q, Tensor! k_cache, Tensor slot_mapping, "
|
||||
"Tensor q_scale_inv, Tensor cache_scale_inv, int cache_block_size, "
|
||||
"Tensor? position_ids=None, Tensor? cos_sin_cache=None) -> ()");
|
||||
ops.def(
|
||||
"fused_kimi_k3_mla_decode_q_concat_ds_mla_insert("
|
||||
"Tensor ql_nope, Tensor q_pe, Tensor kv_c_normed, Tensor k_pe, "
|
||||
"Tensor! mqa_q, Tensor! k_cache, Tensor slot_mapping, "
|
||||
"int cache_block_size, Tensor? position_ids=None, "
|
||||
"Tensor? cos_sin_cache=None) -> ()");
|
||||
|
||||
#ifndef USE_ROCM
|
||||
ops.def(
|
||||
"minimax_allreduce_rms_qk("
|
||||
"Tensor qkv, Tensor norm_weight_q, Tensor norm_weight_k, "
|
||||
"Tensor workspace, int q_size, int kv_size, int rank, int nranks, "
|
||||
"float eps) -> (Tensor, Tensor)");
|
||||
#endif
|
||||
|
||||
#ifndef USE_ROCM
|
||||
ops.def(
|
||||
"fused_qknorm_rope_cache("
|
||||
"Tensor! q_out, Tensor! k_out, Tensor! v_out, Tensor qkv, "
|
||||
"Tensor q_weight, Tensor k_weight, Tensor cos, Tensor sin, "
|
||||
"Tensor positions, Tensor slot_mapping, Tensor! k_cache, "
|
||||
"Tensor! v_cache, int head_dim, int num_q_heads, int num_kv_heads, "
|
||||
"int rotary_pairs, float epsilon, float norm_weight_bias, "
|
||||
"str kv_cache_dtype=\"auto\", Tensor? k_scale=None, Tensor? "
|
||||
"v_scale=None) -> ()");
|
||||
#endif
|
||||
|
||||
ops.def(
|
||||
"fused_muse_spark_qknorm_rope_kv_insert("
|
||||
"Tensor! qkv, Tensor positions, int num_heads, int num_kv_heads, "
|
||||
"int head_dim, float eps, Tensor? cos_sin_cache, int rotary_dim, "
|
||||
"Tensor? slot_mapping, Tensor!? kv_cache, int block_size, "
|
||||
"str kv_cache_dtype, Tensor? k_scale=None, Tensor? v_scale=None, "
|
||||
"Tensor!? q_fp8_out=None, Tensor? q_scale=None) "
|
||||
"-> ()");
|
||||
|
||||
// Horizontally-fused MiniMax-M3 QK-norm + partial NeoX RoPE + KV-insert.
|
||||
ops.def(
|
||||
"fused_minimax_m3_qknorm_rope_kv_insert("
|
||||
"Tensor! qkv, Tensor q_norm_weight, Tensor k_norm_weight, "
|
||||
"Tensor cos_sin_cache, Tensor positions, int num_heads, "
|
||||
"int num_kv_heads, int rotary_dim, float eps, "
|
||||
"Tensor? index_q_norm_weight, Tensor? index_k_norm_weight, "
|
||||
"int num_index_heads, "
|
||||
"Tensor? slot_mapping, Tensor? index_slot_mapping, "
|
||||
"Tensor!? kv_cache, Tensor!? index_cache, "
|
||||
"int block_size, Tensor!? q_out, Tensor!? index_q_out, "
|
||||
"str kv_cache_dtype, bool skip_index_branch=False, "
|
||||
"Tensor!? q_fp8_out=None, float q_fp8_scale=1.0, "
|
||||
"Tensor? kv_k_scale=None, Tensor? kv_v_scale=None) -> ()");
|
||||
|
||||
#ifdef VLLM_ENABLE_FUSED_KDA_DECODE
|
||||
ops.def(
|
||||
"fused_kda_decode("
|
||||
@@ -577,16 +444,6 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
|
||||
"Tensor? state_indices=None, Tensor? has_initial_state=None) -> ()");
|
||||
#endif
|
||||
|
||||
#ifdef VLLM_ENABLE_KIMI_K3_ATTN_RES
|
||||
ops.def(
|
||||
"kimi_k3_attn_res("
|
||||
"Tensor! prefix, Tensor? delta, Tensor! blocks, Tensor norm_weight, "
|
||||
"Tensor qk_weight, Tensor? output_norm_weight, Tensor! output, "
|
||||
"int num_blocks, "
|
||||
"int block_write_idx, float eps, "
|
||||
"float output_norm_eps) -> ()");
|
||||
#endif
|
||||
|
||||
// Apply repetition penalties to logits in-place.
|
||||
ops.def(
|
||||
"apply_repetition_penalties_(Tensor! logits, Tensor prompt_mask, "
|
||||
@@ -619,10 +476,6 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
|
||||
"y_q, Tensor! y_s, bool use_ue8m0) -> ()");
|
||||
ops.def("weak_ref_tensor(Tensor input) -> Tensor");
|
||||
|
||||
ops.def(
|
||||
"muse_spark_mul_tanh_soft_cap(Tensor! out, Tensor input, "
|
||||
"float multiplier, float soft_cap_temp) -> ()");
|
||||
|
||||
// Activation function used in SwiGLU.
|
||||
ops.def("silu_and_mul(Tensor! result, Tensor input) -> ()");
|
||||
|
||||
@@ -799,9 +652,6 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
|
||||
// AWQ ops
|
||||
ops.impl("awq_gemm", TORCH_BOX(&awq_gemm));
|
||||
ops.impl("awq_dequantize", TORCH_BOX(&awq_dequantize));
|
||||
|
||||
// DSV3 fused A GEMM: conditionally compiled so impl registration is in
|
||||
// source file (dsv3_fused_a_gemm.cu)
|
||||
#endif
|
||||
|
||||
ops.impl("merge_attn_states", TORCH_BOX(&merge_attn_states));
|
||||
@@ -825,41 +675,6 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
|
||||
// Positional encoding kernels (shared CUDA/ROCm)
|
||||
ops.impl("rotary_embedding", TORCH_BOX(&rotary_embedding));
|
||||
ops.impl("fused_qk_norm_rope", TORCH_BOX(&fused_qk_norm_rope));
|
||||
ops.impl("fused_deepseek_v4_kv_rope_insert",
|
||||
TORCH_BOX(&fused_deepseek_v4_kv_rope_insert));
|
||||
ops.impl("fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert",
|
||||
TORCH_BOX(&fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert));
|
||||
ops.impl(
|
||||
"fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert",
|
||||
TORCH_BOX(&fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert));
|
||||
ops.impl(
|
||||
"fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert",
|
||||
TORCH_BOX(&fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert));
|
||||
ops.impl("fused_kimi_k3_mla_key_concat_kv_cache_insert",
|
||||
TORCH_BOX(&fused_kimi_k3_mla_key_concat_kv_cache_insert));
|
||||
ops.impl("fused_kimi_k3_mla_key_concat_ds_mla_insert",
|
||||
TORCH_BOX(&fused_kimi_k3_mla_key_concat_ds_mla_insert));
|
||||
ops.impl("fused_kimi_k3_mla_kv_concat",
|
||||
TORCH_BOX(&fused_kimi_k3_mla_kv_concat));
|
||||
ops.impl("fused_kimi_k3_mla_kv_concat_quant_fp8",
|
||||
TORCH_BOX(&fused_kimi_k3_mla_kv_concat_quant_fp8));
|
||||
ops.impl("fused_kimi_k3_mla_qkv_quant_kv_cache_fp8_insert",
|
||||
TORCH_BOX(&fused_kimi_k3_mla_qkv_quant_kv_cache_fp8_insert));
|
||||
ops.impl("fused_kimi_k3_mla_decode_q_concat_kv_cache_insert",
|
||||
TORCH_BOX(&fused_kimi_k3_mla_decode_q_concat_kv_cache_insert));
|
||||
ops.impl("fused_kimi_k3_mla_decode_q_concat_kv_cache_fp8_insert",
|
||||
TORCH_BOX(&fused_kimi_k3_mla_decode_q_concat_kv_cache_fp8_insert));
|
||||
ops.impl("fused_kimi_k3_mla_decode_q_concat_ds_mla_insert",
|
||||
TORCH_BOX(&fused_kimi_k3_mla_decode_q_concat_ds_mla_insert));
|
||||
#ifndef USE_ROCM
|
||||
ops.impl("minimax_allreduce_rms_qk", TORCH_BOX(&minimax_allreduce_rms_qk));
|
||||
ops.impl("fused_qknorm_rope_cache",
|
||||
TORCH_BOX(&vllm::fused_qknorm_rope_cache));
|
||||
#endif
|
||||
ops.impl("fused_muse_spark_qknorm_rope_kv_insert",
|
||||
TORCH_BOX(&fused_muse_spark_qknorm_rope_kv_insert));
|
||||
ops.impl("fused_minimax_m3_qknorm_rope_kv_insert",
|
||||
TORCH_BOX(&fused_minimax_m3_qknorm_rope_kv_insert));
|
||||
#ifdef VLLM_ENABLE_FUSED_KDA_DECODE
|
||||
ops.impl("fused_kda_decode", TORCH_BOX(&fused_kda_decode));
|
||||
#endif
|
||||
@@ -874,10 +689,6 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
|
||||
ops.impl("fused_kda_chunk", TORCH_BOX(&fused_kda_chunk));
|
||||
#endif
|
||||
|
||||
#ifdef VLLM_ENABLE_KIMI_K3_ATTN_RES
|
||||
ops.impl("kimi_k3_attn_res", TORCH_BOX(&kimi_k3_attn_res));
|
||||
#endif
|
||||
|
||||
// Sampler kernels (shared CUDA/ROCm)
|
||||
ops.impl("apply_repetition_penalties_",
|
||||
TORCH_BOX(&apply_repetition_penalties_));
|
||||
@@ -893,8 +704,6 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) {
|
||||
TORCH_BOX(&persistent_masked_m_silu_mul_quant));
|
||||
ops.impl("weak_ref_tensor", TORCH_BOX(&weak_ref_tensor));
|
||||
ops.impl("silu_and_mul_quant", TORCH_BOX(&silu_and_mul_quant));
|
||||
ops.impl("muse_spark_mul_tanh_soft_cap",
|
||||
TORCH_BOX(&muse_spark_mul_tanh_soft_cap));
|
||||
ops.impl("silu_and_mul", TORCH_BOX(&silu_and_mul));
|
||||
ops.impl("mul_and_silu", TORCH_BOX(&mul_and_silu));
|
||||
ops.impl("gelu_and_mul", TORCH_BOX(&gelu_and_mul));
|
||||
|
||||
Reference in New Issue
Block a user