ggml : add NVFP4 quantization type support (#19769)
* WIP: add NVFP4 quantization support * tests * improve NVFP4 dot product implementation performance and fix bad super call * typo * Use nvfp4 kvalues * vulkan : fix NVFP4 shader compilation by including kvalues_mxfp4 lookup table * vulcal and perf fixes * wip * Fix metal * fix vulcan * Rename threshold & fix wrong scale * Fix MOE * Shelf backend implementations (CUDA, Metal, Vulkan, arch-specific SIMD) Remove NVFP4 support from GPU backends and architecture-specific optimized dot products. These should be added in separate PRs so backend specialists can review them independently. Reverted files: - ggml-cuda: common.cuh, convert.cu, mmq.cu/cuh, mmvq.cu, vecdotq.cuh, quantize.cu/cuh, mma.cuh, ggml-cuda.cu, fattn-tile.cuh - ggml-metal: ggml-metal.metal, ggml-metal-device.cpp, ggml-metal-impl.h, ggml-metal-ops.cpp - ggml-vulkan: ggml-vulkan.cpp, all vulkan-shaders/* - ggml-cpu arch: arm/quants.c, x86/quants.c, powerpc/quants.c, s390/quants.c Core NVFP4 support (type definition, CPU fallback dot product, quantization, dequantization, conversion) is retained. * Fix arch-fallback.h: add NVFP4 generic fallback for all platforms After shelving backend-specific SIMD implementations, the generic CPU dot product needs to be aliased on ARM, x86, PowerPC, and s390 platforms that previously relied on arch-specific versions. * quantize: add NVFP4 as a quantization type option * Fix ggml_fp32_to_ue4m3: handle subnormal values Previously, values with ue4m3_exp <= 0 were clamped to 0, causing all small scales to underflow. This made NVFP4 quantization via llama-quantize produce garbage (PPL = 5.8M) since typical transformer weights have amax/6.0 in the range 0.001-0.01, which falls in the UE4M3 subnormal range. Now subnormals are properly encoded as man * 2^-9 (exp=0, man=1..7), matching the decode path in ggml_ue4m3_to_fp32. Result: NVFP4 requantization now produces PPL = 15.25 (vs F16 = 14.33), comparable to Q4_1 (PPL = 15.81) at slightly lower BPW (4.70 vs 5.15). * Restore ARM NEON NVFP4 dot product implementation Restores the optimized ggml_vec_dot_nvfp4_q8_0 for ARM NEON using vqtbl1q_s8 lookup and ggml_vdotq_s32 dot products. tg128 performance: 4.37 t/s (generic) -> 13.66 t/s (NEON) = 3.1x speedup * Optimize ARM NEON NVFP4 dot product: LUT + vpaddq + vfmaq - Add ue4m3_scale_lut[128] to ggml-common.h replacing branch-heavy ggml_ue4m3_to_fp32() in the hot loop - Use vpaddq_s32 for pairwise int32 reduction instead of vaddvq_s32 - Accumulate with vfmaq_f32 into float32x4_t vector accumulators tg128: 8.1 -> 31.0 t/s (3.8x speedup, 77% of Q4_1 speed) * ARM NEON NVFP4: rearrange q8 to match nibble layout Alternative approach: rearrange q8 data to match the NVFP4 lo/hi nibble layout instead of rearranging the looked-up NVFP4 values. Eliminates vcombine_s8(vget_low, vget_low) shuffles. Performance is equivalent (~18.5 t/s) - the bottleneck is the 2x block overhead from QK=16 vs QK=32, not the shuffle instructions. * CPU only backend 64 super-block layout * cleanup * Remove unused LUT * int * exclude NVFP4 from unsupported ops in metal build * remove quantization for now * store scales as native UE4M3, preserve original model bits when possible * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * correct comment * format * reduce duplication and cleanup * Address comments * move detection to prepare_tensors * Use math instead of const * Move * fix comment * Shelf quantize tests * Rebase and move check * cleanup * lint * Update gguf-py/gguf/scripts/gguf_convert_endian.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Use fallback quant config * Simplify Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * organize * Refactor * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update convert_hf_to_gguf.py Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * add quantize_nvfp4 (required for test_quants.py) * add quantize_nvfp4 (required for test_quants.py) * add quantize_nvfp4 (required for test_quants.py) * fix return type --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
Richard Davison
GitHub
parent
3ca19b0e9f
commit
5eae9cb1d9
31 changed files with 710 additions and 51 deletions
|
|
@ -30,8 +30,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
|
|||
{
|
||||
// compute Q and K and RoPE them
|
||||
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
|
||||
if (model.layers[il].wq_scale) {
|
||||
Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale);
|
||||
if (model.layers[il].wq_s) {
|
||||
Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_s);
|
||||
}
|
||||
cb(Qcur, "Qcur", il);
|
||||
if (model.layers[il].bq) {
|
||||
|
|
@ -41,8 +41,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
|
|||
|
||||
// B1.K
|
||||
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
|
||||
if (model.layers[il].wk_scale) {
|
||||
Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale);
|
||||
if (model.layers[il].wk_s) {
|
||||
Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_s);
|
||||
}
|
||||
cb(Kcur, "Kcur", il);
|
||||
if (model.layers[il].bk) {
|
||||
|
|
@ -52,8 +52,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
|
|||
|
||||
// B1.V
|
||||
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
|
||||
if (model.layers[il].wv_scale) {
|
||||
Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale);
|
||||
if (model.layers[il].wv_s) {
|
||||
Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_s);
|
||||
}
|
||||
cb(Vcur, "Vcur", il);
|
||||
if (model.layers[il].bv) {
|
||||
|
|
@ -91,8 +91,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
|
|||
cb(cur, "attn_sub_norm", il);
|
||||
|
||||
cur = build_lora_mm(model.layers[il].wo, cur);
|
||||
if (model.layers[il].wo_scale) {
|
||||
cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale);
|
||||
if (model.layers[il].wo_s) {
|
||||
cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);
|
||||
}
|
||||
if (model.layers[il].bo) {
|
||||
cur = ggml_add(ctx0, cur, model.layers[il].bo);
|
||||
|
|
@ -115,8 +115,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
|
|||
cb(cur, "ffn_norm", il);
|
||||
|
||||
cur = build_ffn(cur,
|
||||
model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_scale,
|
||||
model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale,
|
||||
model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
|
||||
model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
|
||||
NULL, NULL, NULL,
|
||||
NULL,
|
||||
LLM_FFN_SILU, LLM_FFN_PAR, il);
|
||||
|
|
@ -128,8 +128,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
|
|||
cb(cur, "ffn_sub_norm", il);
|
||||
|
||||
cur = build_lora_mm(model.layers[il].ffn_down, cur);
|
||||
if (model.layers[il].ffn_down_scale) {
|
||||
cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale);
|
||||
if (model.layers[il].ffn_down_s) {
|
||||
cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_s);
|
||||
}
|
||||
cb(cur, "ffn_down", il);
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue