From 55c1b2a3bbd470e9e2a3a0618b92cf64a885f806 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Tue, 26 Mar 2024 15:21:27 +0100 Subject: IQ1_M: 1.75 bpw quantization (#6302) * iq1_m: basics * iq1_m: basics-2 * iq1_m: CUDA dequantize works Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B. * iq1_m: separate shifts for each group of 8 in a block We get PPL(LLaMA-v2-7B ) = 9.2810 PPL(LLaMA-v2-13B) = 6.8105 Not bad, but slightly higher than sqrt(PPL(IQ1_S) * PPL(IQ2_XXS)) which is the expected outcome given that IQ1_M is halfway between IQ1_S and IQ2_XXS in terms of bpw. From this, we would expect PPL = 9.14 for LLaMA-v2-7B PPL = 6.63 for LLaMA-v2-13B * iq1_m: go to 3-bit scales There is slight increase in PPL, but the 0.0625 bpw reduction in size is totally worth it. We now have PPL(LLaMA-v2-7B ) = 9.4469 at 1.96 bpw PPL(LLaMA-v2-13B) = 6.8717 at 1.93 bpw PPL(LLaMA-v2-70B) = 4.8568 at 1.85 bpw * iq1_m: scalar dot product * iq1_m: AVX2 dot product * iq1_m: very slightly faster AVX2 dot product * iq1_m: ARM_NEON dot product Works, but very slow (10.5 t/s) * iq1_m: Metal - dequantize works, dot product does not * iq1_m: Metal now works About the same performance as iq1_s. * iq1_m: minor * iq1_m: checking pure iq1_m quantization It is pretty bad: PPL(LLaMA-v2-7B) = 34 if we quantize output.weight with Q4_K. * iiq1_m: slightly faster ARM_NEON dot product 10.5 t/s -> 11.65 t/s * iq1_m: faster ARM_NEON dot product 11.65 t/s -> 14.9 t/s * iq1_m: another minor ARM_NEON dot product improvement 14.9 -> 15.0 t/s * iq1_m: small PPL improvement via super-block scale adjustment After quantizing block scales redo the super-block scale fit. PPL(LLaMA-v2-7B ) = 9.3346 PPL(LLaMA-v2-13B) = 6.8419 PPL(LLaMA-v2-70B) = 4.8294 PPL(Mistral-7B ) = 8.1624 * iq1_m: adapt to CUDA refactoring * iq1_m: remove unused variable We have progressed to warnings being errors. * iq1_m: add to backend-ops tests * iq1_m: fix Windows ARM * iq1_m: use common definition of iq1m_scale_t * cuda: assert -> NO_DEVICE_CODE * iq1_M: PR comments --------- Co-authored-by: Iwan Kawrakow --- llama.cpp | 24 +++++++++++++++++++----- 1 file changed, 19 insertions(+), 5 deletions(-) (limited to 'llama.cpp') diff --git a/llama.cpp b/llama.cpp index 6980311b..68c360c7 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3018,6 +3018,7 @@ struct llama_model_loader { case GGML_TYPE_IQ2_S: ftype = LLAMA_FTYPE_MOSTLY_IQ2_S; break; case GGML_TYPE_IQ3_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ3_XXS; break; case GGML_TYPE_IQ1_S: ftype = LLAMA_FTYPE_MOSTLY_IQ1_S; break; + case GGML_TYPE_IQ1_M: ftype = LLAMA_FTYPE_MOSTLY_IQ1_M; break; case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break; case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break; case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break; @@ -3413,6 +3414,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { 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_IQ1_S :return "IQ1_S - 1.5625 bpw"; + case LLAMA_FTYPE_MOSTLY_IQ1_M :return "IQ1_M - 1.75 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_NL: return "IQ4_NL - 4.5 bpw"; case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw"; @@ -12447,7 +12449,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n new_type = GGML_TYPE_Q8_0; } 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_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || + ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) { new_type = GGML_TYPE_Q5_K; } else if (new_type != GGML_TYPE_Q8_0) { @@ -12458,7 +12461,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n if (qs.params->token_embedding_type < GGML_TYPE_COUNT) { new_type = qs.params->token_embedding_type; } else { - if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) { + if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || + ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) { new_type = GGML_TYPE_Q2_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) { @@ -12469,7 +12473,7 @@ 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_IQ1_S || - ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) { + ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) { if (name.find("attn_v.weight") != std::string::npos) { if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q4_K; else new_type = ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M ? GGML_TYPE_IQ3_S : GGML_TYPE_Q2_K; @@ -12488,7 +12492,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n if (qs.model.hparams.n_expert == 8) { new_type = GGML_TYPE_Q5_K; } else { - if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) new_type = GGML_TYPE_IQ2_XXS; + if (ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) new_type = GGML_TYPE_IQ2_XXS; else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M) new_type = GGML_TYPE_IQ3_S; } } @@ -12655,7 +12659,8 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K || new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K || new_type == GGML_TYPE_IQ4_XS || new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_S || - new_type == GGML_TYPE_IQ3_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S || new_type == GGML_TYPE_IQ3_S) { + new_type == GGML_TYPE_IQ3_XXS || new_type == GGML_TYPE_IQ1_S || new_type == GGML_TYPE_IQ3_S || + new_type == GGML_TYPE_IQ1_M) { int nx = tensor->ne[0]; int ny = tensor->ne[1]; if (nx % QK_K != 0) { @@ -12673,6 +12678,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n case GGML_TYPE_IQ3_XXS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: case GGML_TYPE_Q2_K: case GGML_TYPE_Q3_K: case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break; @@ -12754,6 +12760,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ2_M: default_type = GGML_TYPE_IQ2_S; break; case LLAMA_FTYPE_MOSTLY_IQ3_XXS: default_type = GGML_TYPE_IQ3_XXS; break; case LLAMA_FTYPE_MOSTLY_IQ1_S: default_type = GGML_TYPE_IQ1_S; break; + case LLAMA_FTYPE_MOSTLY_IQ1_M: default_type = GGML_TYPE_IQ1_M; break; case LLAMA_FTYPE_MOSTLY_IQ4_NL: default_type = GGML_TYPE_IQ4_NL; break; case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break; case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break; @@ -12929,6 +12936,12 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s if (!params->pure && ggml_is_quantized(default_type)) { new_type = llama_tensor_get_type(qs, new_type, tensor, ftype); } + else if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) { + new_type = params->token_embedding_type; + } + else if (params->output_tensor_type < GGML_TYPE_COUNT && strcmp(tensor->name, "output.weight") == 0) { + new_type = params->output_tensor_type; + } // If we've decided to quantize to the same type the tensor is already // in then there's nothing to do. @@ -12961,6 +12974,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_S || new_type == GGML_TYPE_IQ1_S || + (new_type == GGML_TYPE_IQ1_M && strcmp(tensor->name, "token_embd.weight") && strcmp(tensor->name, "output.weight")) || (new_type == GGML_TYPE_Q2_K && params->ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(tensor->name, "token_embd.weight") != 0)) && !imatrix) { LLAMA_LOG_ERROR("\n\n============================================================\n"); LLAMA_LOG_ERROR("Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name); -- cgit v1.2.3