From a1c931c30ce9c5618ec56fe93234110343111710 Mon Sep 17 00:00:00 2001 From: Andrew Chan Date: Thu, 22 May 2025 23:17:52 -0700 Subject: Trellis quants with CPU inference (#441) * WIP * WIP * WIP * Testing Trellis quantization Using 12 bits per 8 weights I get a better rmse than iq2_xxs. I still need to see how quantizing the group-of-8 scales will affect accuracy. By AVX2 SIMDifying the search for the best code, LLaMA-3.1-8B gets quantized in 130 seconds on the Ryzen-7950X CPU - sluggish but still acceptable. * Testing Trellis quantization: 4-bit quantized block scales rmse increases by just 3%, so this is beating iq2_xss in terms of rmse at the same 2.0625 bpw. * Testing Trellis quantization: playing with scales and generators * iq2_kt: quantize / dequantize I now see that I was comparing apples to oranges: iq2_xxs was using a weight of sigma^2/4 + x^2, while the Trellis approach wasn't (weight = 1). Once I use the same weight, iq2_kt is actually slightly worse than iq2_xxs in terms of rmse, so does not look promising at this point. Also, once each group of 8 Trellis values no longer has a constant sum(q^2) that we can precompute, quantization becomes significantly slower (476 seconds for LLaMA-3.1-8B). * iq2_kt: CUDA dequantize so we can run perplexity calcs. As already indicated by rmse, the 2-bit trellis approach is quite a bit worse than iq2_xxs. * WIP * WIP * WIP - try larger blocks With blocks of 32 and 16 bits per groups of 8 the brute force seach becomes prohibitive in terms of CPU time (30+ minutes for 8B LLaMA after SIMDifying with AVX2). The trick is to group the points in clusters, find the nearest cluster, and only search within the cluster. * iq2_kt - this is better Using blocks of 32 and 16 bits per group of 8 weights it beats iq2_xxs in terms of PPL by a significant margin. It is 0.0625 bpw larger, but even if we go to 15 bits per group od 8 (so 0.0625 bpw less than iq2_xxs), PPL is still lower. * iq2_kt - even better Re-quantize after determining block scales (at the epxense of much longer quantization time). * iq2_kt: CUDA dot product Implemented as DMMV. Very slow - just 81 t/s for LLaMA-3.1-8B. Then again, Q2_K_S with forced to use DMMV only gets 112 t/s vs 145 t/s via MMVQ. My memory is that when the DMMV kernels were properly maintained/used, DMMV was about on par with MMVQ for k-quants on my GPU. * iq2_kt: very slightly faster CUDA dot product * iq2_kt: f16 CUDA dot product We arrive at 112 t/s. * iq2_kt: faster f16 CUDA dot product We arrive at 139 t/s (no FA), and 149 t/s (FA). My RTX-4080 is ~20% slower than the RTX-6000 quoted in the QTIP repository, so with FA (which I'm sure they also used) we are at around ~180 t/s on their GPU, so almost matching their performance. * iq2_kt: faster f16 CUDA dot product We arrive at 146 t/s (no FA), and 158 t/s (FA). This is measured for LLaMA-3.1-8B with output.weight left as f16. * Minor * Adding iq3_kt 3.125 bpw. So far does not look good on the PPL vs bpw plot. * Forgotten change * WIP * WIP * iq3_kt WIP: slowly improving PPL(LLaMA-3.1-8B-Instruct, 8192) is now 6.8322, which is starting to be competitive/slightly better than other quants. * WIP * iq3_kt WIP: slowly improving PPL(LLaMA-3.1-8B-Instruct, 8192) is now 6.7892 * iq3_kt WIP: slowly improving PPL(LLaMA-3.1-8B-Instruct, 8192) is now 6.7689 after shrinking by 0.015 bpw by using iq4_k instead of q5_k for attn_v. * iq3_kt WIP: speed up quantization Nearly 60% improvement of quantization speed by having the points nelonging to a cluster copied to contiguous memory during initialization, and then accessed sequantially while searching for the closest point. LLaMA-3.1-8B now gets quantized in ~150 seconds on the Ryzen-5975WX. * iq3_kt speed up quantization Same trick as last commit applied to iq2_kt. Here we get an even larger speedup: quantization time on the Ryzen-5975WX for LLaMA-3.1-8B drops to 195 seconds from 375 seconds! * iq3_kt: CUDA dot product * iq2_kt: SOTA We arrive at PPL(LLaMA-3.1-8B-Instruct, 8192) = 9.2406 PPL(LLaMA-2-7B, 4096) = 6.4179 * iq2_kt: SOTA We arrive at PPL(LLaMA-3.1-8B-Instruct, 8192) = 9.1642 PPL(LLaMA-2-7B, 4096) = 6.3920 * Adding iq4_kt - not competitive at this point * WIP * WIP * iq4_kt: CUDA dot product * iq4_kt: minor tweaks * iq2_kt: SOTA We arrive at PPL(LLaMA-3.1-8B-Instruct, 8192) = 9.1642 PPL(LLaMA-2-7B, 4096) = 6.3920 * iq2_kt: SOTA We arrive at PPL(LLaMA-3.1-8B-Instruct, 8192) = 9.0297 PPL(LLaMA-2-7B, 4096) = 6.3913 Ah, quantization is faster too. About 20% faster. * iq3_kt: small improvements and faster quantization * iq2_kt: SOTA We arrive at PPL(LLaMA-3.1-8B-Instruct, 8192) = 8.9627 PPL(LLaMA-2-7B, 4096) = 6.3825 Quantization is faster too: ~200 seconds for LLaMA-3.1-8B on Ryzen-5975WX. * iq3_kt: small progress * WIP * iq4_kt: go to 4.0 bpw 15 bits per group of 4, plus 8 bit scales ifor blocks of 32. This gives a slightly better PPL than iq4_kss. * iq4_kt: very slightly better at the expense of much longer quantization time. * iq4_kt: failed attemt to adjust CUDA dot product It was working for 4.125 bpw. But after changing to 4.0 bpw there is something wrong and I don't see the bug. * DRY * DRY * iq4_kt: CUDA dot product works * DRY * Report actual bpw * Minor tweaks * Checkpoint Go to groups of 8 for iq3_kt. 2 x 8 = 16 bits for the magnitude plus 1 bpw for the sign. It goves a visible improvement in the PPL vs bpw plot, but that comes at the expense of much longer quantization time (7.5 minutes for LLaMA-3.1-8B on the Ryzen-5975WX). I also notices that the 3INST generator is not actually generating a Gaussian distribution. But going to a better generator means readjusting all the hyper-parameters, so leaving it for later. * WIP for IQ2_KT * WIP - working basic iq2_kt * still super slow (0.17t/s eval) * flatten 3inst iters + avx2 (0.3t/s eval) * iq3_kt (0.3t/s eval) and renames * wip buggy iq4_KT * fix (0.22t/s eval) * naming and remove unused fn * cleanup * more cleanup * delete unused and noncompiling mmvq functions * Some performance tweaks * Slighty faster iq2_kt * port Trellis struct to iq3_kt, iq4_kt * oops untracked files --------- Co-authored-by: Iwan Kawrakow --- src/llama.cpp | 73 ++++++++++++++++++++++++++++++++++++++++++++++++++++++----- 1 file changed, 67 insertions(+), 6 deletions(-) (limited to 'src') diff --git a/src/llama.cpp b/src/llama.cpp index 9d9c7c4e..48d7214d 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -4355,6 +4355,9 @@ struct llama_model_loader { case GGML_TYPE_IQ2_S_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ2_M_R4;break; case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break; case GGML_TYPE_IQ3_XXS_R4: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4; break; + case GGML_TYPE_IQ2_KT: ftype = LLAMA_FTYPE_MOSTLY_IQ2_KT; break; + case GGML_TYPE_IQ3_KT: ftype = LLAMA_FTYPE_MOSTLY_IQ3_KT; break; + case GGML_TYPE_IQ4_KT: ftype = LLAMA_FTYPE_MOSTLY_IQ4_KT; break; case GGML_TYPE_IQ1_S: ftype = LLAMA_FTYPE_MOSTLY_IQ1_S; break; case GGML_TYPE_IQ1_S_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ1_S_R4;break; case GGML_TYPE_IQ1_M_R4:ftype = LLAMA_FTYPE_MOSTLY_IQ1_M_R4;break; @@ -5095,6 +5098,9 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ2_M_R4: return "IQ2_M_R4 - 2.7 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_XS: return "IQ3_XS - 3.3 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_XXS: return "IQ3_XXS - 3.0625 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ2_KT: return "IQ2_KT - 2.125 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ3_KT: return "IQ3_KT - 3.125 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ4_KT: return "IQ4_KT - 4.0 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4: return "IQ3_XXS_R4 - 3.0625 bpw"; case LLAMA_FTYPE_MOSTLY_IQ1_S: return "IQ1_S - 1.5625 bpw"; case LLAMA_FTYPE_MOSTLY_IQ1_S_R4: return "IQ1_S_R4 - 1.5 bpw"; @@ -18787,10 +18793,11 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K || - ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4 || - ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 || - ftype == LLAMA_FTYPE_MOSTLY_IQ2_M_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S_R4 || - ftype == LLAMA_FTYPE_MOSTLY_IQ1_M_R4) { + ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 || + ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 || + ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M_R4 || + ftype == LLAMA_FTYPE_MOSTLY_IQ1_S_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M_R4 || + ftype == LLAMA_FTYPE_MOSTLY_IQ2_KT || ftype == LLAMA_FTYPE_MOSTLY_IQ3_KT) { new_type = !qs.has_output ? GGML_TYPE_IQ4_K : GGML_TYPE_Q5_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS_R4) { @@ -18818,7 +18825,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M_R4) { new_type = GGML_TYPE_IQ3_S; } - else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_KT) { new_type = GGML_TYPE_IQ3_S; } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) { @@ -18863,6 +18870,42 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (name.find("attn_output.weight") != std::string::npos) { new_type = qs.model.hparams.n_expert >= 4 ? GGML_TYPE_Q5_K_R4 : GGML_TYPE_IQ2_K_R4; } + } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_KT) { + if (name.find("attn_v.weight") != std::string::npos) { + if (qs.model.hparams.n_expert >= 4 || qs.model.hparams.n_gqa() >= 4) new_type = GGML_TYPE_IQ4_K; + else if (qs.model.hparams.n_gqa() >= 2) new_type = GGML_TYPE_IQ3_K; + else new_type = GGML_TYPE_Q2_K; + ++qs.i_attention_wv; + } + else if (qs.model.hparams.n_expert >= 8 && name.find("attn_k") != std::string::npos) { + new_type = GGML_TYPE_Q4_K; + } + else if (qs.model.hparams.n_expert >= 8 && (name.find("blk.0.ffn_down") != std::string::npos || + name.find("blk.0.ffn_gate") != std::string::npos || + name.find("blk.0.ffn_up") != std::string::npos)) { + new_type = GGML_TYPE_IQ3_K; + } + else if (qs.model.hparams.n_expert >= 8 && name.find("attn_q") != std::string::npos) { + new_type = GGML_TYPE_Q4_K; + } + else if (name.find("attn_qkv.weight") != std::string::npos) { + new_type = GGML_TYPE_IQ3_K; + } + else if (name.find("_shexp.weight") != std::string::npos) { + new_type = GGML_TYPE_IQ4_K; + } + else if (name.find("ffn_down") != std::string::npos) { + auto [i_layer, n_layer] = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str()); + if (qs.params->ffn_down_type < GGML_TYPE_COUNT) new_type = qs.params->ffn_down_type; + else if (i_layer < n_layer/8) { + new_type = GGML_TYPE_IQ3_K; + } + ++qs.i_ffn_down; + } + else if (name.find("attn_output.weight") != std::string::npos) { + new_type = qs.model.hparams.n_expert >= 4 ? GGML_TYPE_Q5_K : GGML_TYPE_IQ3_K; + } } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M || ftype == LLAMA_FTYPE_MOSTLY_IQ2_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS_R4 || @@ -18919,6 +18962,16 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ3_K : !qs.has_imatrix ? GGML_TYPE_IQ3_S : GGML_TYPE_IQ3_XXS; } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KT) { + //new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_IQ4_K : qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ3_K + // : !qs.has_imatrix ? GGML_TYPE_IQ3_K : GGML_TYPE_IQ3_KT; + new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_IQ4_K : GGML_TYPE_IQ3_K; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ4_KT) { + //new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_IQ5_K : qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ4_K + // : !qs.has_imatrix ? GGML_TYPE_IQ4_KS : GGML_TYPE_IQ4_KT; + new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_IQ5_K : GGML_TYPE_IQ4_K; + } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) { new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K_R4 : qs.model.hparams.n_gqa() >= 2 ? GGML_TYPE_IQ3_K_R4 : !qs.has_imatrix ? GGML_TYPE_IQ3_K_R4 : GGML_TYPE_IQ3_XXS_R4; @@ -19046,6 +19099,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && !qs.has_imatrix) { new_type = i_layer < n_layer/8 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K; } + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KT && !qs.has_imatrix) { + new_type = i_layer < n_layer/8 ? GGML_TYPE_IQ4_K : GGML_TYPE_IQ3_K; + } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 && !qs.has_imatrix) { new_type = i_layer < n_layer/8 ? GGML_TYPE_Q4_K_R4 : GGML_TYPE_IQ3_K_R4; } @@ -19110,7 +19166,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n ftype == LLAMA_FTYPE_MOSTLY_IQ4_KSS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ4_KS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ5_KS || ftype == LLAMA_FTYPE_MOSTLY_IQ5_KS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_R4 || - ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R8 || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_R4 || + ftype == LLAMA_FTYPE_MOSTLY_IQ4_NL_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ4_XS_R8 || + ftype == LLAMA_FTYPE_MOSTLY_Q3_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_KT || ftype == LLAMA_FTYPE_MOSTLY_Q2_K_R4|| ftype == LLAMA_FTYPE_MOSTLY_IQ4_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ2_K_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4 || ftype == LLAMA_FTYPE_MOSTLY_IQ3_S_R4) { new_type = GGML_TYPE_Q5_K; // should the IQ_K quants be applied here as the new type for the IQ_K ftypes ? @@ -19119,6 +19176,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n } else { if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K; // This list could be generalized and streamlined else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) new_type = GGML_TYPE_IQ3_S; + else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_KT && qs.model.hparams.n_gqa() >= 4) new_type = GGML_TYPE_IQ3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4) new_type = GGML_TYPE_IQ3_K_R4; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ) new_type = GGML_TYPE_Q4_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L ) new_type = GGML_TYPE_Q5_K; @@ -19321,10 +19379,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ2_XS: default_type = GGML_TYPE_IQ2_XS; break; case LLAMA_FTYPE_MOSTLY_IQ2_XS_R4:default_type = GGML_TYPE_IQ2_XS_R4; break; case LLAMA_FTYPE_MOSTLY_IQ2_KS: default_type = GGML_TYPE_IQ2_KS; break; + case LLAMA_FTYPE_MOSTLY_IQ2_KT: default_type = GGML_TYPE_IQ2_KT; break; case LLAMA_FTYPE_MOSTLY_IQ2_S: default_type = GGML_TYPE_IQ2_XS; break; case LLAMA_FTYPE_MOSTLY_IQ2_M: default_type = GGML_TYPE_IQ2_S; break; case LLAMA_FTYPE_MOSTLY_IQ2_M_R4:default_type = GGML_TYPE_IQ2_S_R4;break; case LLAMA_FTYPE_MOSTLY_IQ3_XXS: default_type = GGML_TYPE_IQ3_XXS; break; + case LLAMA_FTYPE_MOSTLY_IQ3_KT: default_type = GGML_TYPE_IQ3_KT; break; + case LLAMA_FTYPE_MOSTLY_IQ4_KT: default_type = GGML_TYPE_IQ4_KT; break; case LLAMA_FTYPE_MOSTLY_IQ3_XXS_R4: default_type = GGML_TYPE_IQ3_XXS_R4; break; case LLAMA_FTYPE_MOSTLY_IQ1_S: default_type = GGML_TYPE_IQ1_S; break; case LLAMA_FTYPE_MOSTLY_IQ1_S_R4:default_type = GGML_TYPE_IQ1_S_R4;break; -- cgit v1.2.3