From ce530b0fd0a64bf815c614fc6b59b4e164910557 Mon Sep 17 00:00:00 2001 From: isHuangXin Date: Mon, 25 May 2026 03:01:27 +0800 Subject: [PATCH] Add bitnet-embeddings-270m model adaptation with F16 and I2_S GGUF conversion - Add LLM_ARCH_GEMMA3 in llama.cpp for gemma3_text model type (embedding scaling, GELU, post-attn/post-FFN norms, GQA) - Add GGUF conversion support for Gemma3-based 270m models (SPM tokenizer, RMSNorm w+1 offset, arch-specific tensor mapping) - Add tokenizer hash for multilingual-e5-0.6b-260311 - Add conversion documentation --- docs/bitnet-embeddings-gguf-conversion.md | 410 ++++++++++++++++++ ...bitnet-embeddings-qwen3-gguf-conversion.md | 302 ------------- utils/convert-bitnet-embedding-to-gguf.py | 234 ++++++++-- 3 files changed, 599 insertions(+), 347 deletions(-) create mode 100644 docs/bitnet-embeddings-gguf-conversion.md delete mode 100644 docs/bitnet-embeddings-qwen3-gguf-conversion.md diff --git a/docs/bitnet-embeddings-gguf-conversion.md b/docs/bitnet-embeddings-gguf-conversion.md new file mode 100644 index 0000000..a4ee919 --- /dev/null +++ b/docs/bitnet-embeddings-gguf-conversion.md @@ -0,0 +1,410 @@ +# BitNet Embeddings GGUF Conversion Implementation + +## 1. Background + +BitNet embedding models apply per-projection RMSNorm (`BitLinear`) before each linear projection (q/k/v/o/gate/up/down). Each projection has a `.norm.weight` that applies RMSNorm to the input **before** the matmul: + +``` +x → RMSNorm(x, norm.weight) → activation_quant(8bit) → matmul(weight_quant(ternary)) +``` + +This pattern does **not** exist in any standard llama.cpp architecture: +- Standard Qwen3/Gemma3: no per-projection norms +- Standard BitNet: has `attn_sub_norm`/`ffn_sub_norm` at different positions (after attention/gate*up, not before each projection) + +Currently two base architectures are supported: + +| | bitnet-embeddings-0.6b (Qwen3) | bitnet-embeddings-270m (Gemma3) | +|---|---|---| +| Architecture | `Qwen3Model` | `Gemma3TextModel` | +| hidden_size | 1024 | 640 | +| num_attention_heads | 16 | 4 | +| num_key_value_heads | 8 | 1 | +| head_dim | 128 (note: != hidden_size/num_heads = 64) | 256 (note: != hidden_size/num_heads = 160) | +| intermediate_size | 3072 | 2048 | +| num_hidden_layers | 28 | 18 | +| hidden_activation | SiLU | gelu_pytorch_tanh | +| vocab_size | 151936 | 262144 | +| rope_theta | 1000000 | 10000.0 | +| rms_norm_eps | 1e-06 | 1e-06 | +| query_pre_attn_scalar | N/A | 256 | +| tie_word_embeddings | true | true | + +### Gemma3 vs Qwen3 Key Differences + +| Feature | Qwen3 | Gemma3 | +|---------|-------|--------| +| Post-attn norm | No | Yes (`post_attention_norm`) | +| Post-FFW norm | No | Yes (`post_ffw_norm`) | +| Pre-FFW norm naming | `post_attention_layernorm` → `ffn_norm` | `pre_feedforward_layernorm` → `ffn_norm` | +| QK head norms | Yes | Yes | +| Activation | SiLU | GELU | +| Embedding scaling | No | sqrt(n_embd) | +| EOS token override | Yes (`<\|endoftext\|>` 151643) | No (auto from tokenizer) | + +### Per-Layer Tensors (7 extra norm tensors per layer) + +| Tensor | Qwen3 Shape | Gemma3 Shape | +|--------|-------------|--------------| +| `self_attn.q_proj.norm.weight` | [1024] | [640] | +| `self_attn.k_proj.norm.weight` | [1024] | [640] | +| `self_attn.v_proj.norm.weight` | [1024] | [640] | +| `self_attn.o_proj.norm.weight` | [2048] | [1024] | +| `mlp.gate_proj.norm.weight` | [1024] | [640] | +| `mlp.up_proj.norm.weight` | [1024] | [640] | +| `mlp.down_proj.norm.weight` | [3072] | [2048] | + +--- + +## 2. GGUF Tensor Name Mapping + +### Common Tensors (both architectures) + +| HF Name | GGUF Name | Notes | +|----------|-----------|-------| +| `embed_tokens.weight` | `token_embd.weight` | | +| `norm.weight` | `output_norm.weight` | | +| `layers.{i}.input_layernorm.weight` | `blk.{i}.attn_norm.weight` | | +| `layers.{i}.self_attn.q_proj.weight` | `blk.{i}.attn_q.weight` | | +| `layers.{i}.self_attn.k_proj.weight` | `blk.{i}.attn_k.weight` | | +| `layers.{i}.self_attn.v_proj.weight` | `blk.{i}.attn_v.weight` | | +| `layers.{i}.self_attn.o_proj.weight` | `blk.{i}.attn_output.weight` | | +| `layers.{i}.self_attn.q_norm.weight` | `blk.{i}.attn_q_norm.weight` | QK head norm | +| `layers.{i}.self_attn.k_norm.weight` | `blk.{i}.attn_k_norm.weight` | QK head norm | +| `layers.{i}.self_attn.q_proj.norm.weight` | `blk.{i}.attn_q_norm_in.weight` | BitNet per-projection | +| `layers.{i}.self_attn.k_proj.norm.weight` | `blk.{i}.attn_k_norm_in.weight` | BitNet per-projection | +| `layers.{i}.self_attn.v_proj.norm.weight` | `blk.{i}.attn_v_norm_in.weight` | BitNet per-projection | +| `layers.{i}.self_attn.o_proj.norm.weight` | `blk.{i}.attn_output_norm_in.weight` | BitNet per-projection | +| `layers.{i}.mlp.gate_proj.weight` | `blk.{i}.ffn_gate.weight` | | +| `layers.{i}.mlp.up_proj.weight` | `blk.{i}.ffn_up.weight` | | +| `layers.{i}.mlp.down_proj.weight` | `blk.{i}.ffn_down.weight` | | +| `layers.{i}.mlp.gate_proj.norm.weight` | `blk.{i}.ffn_gate_norm_in.weight` | BitNet per-projection | +| `layers.{i}.mlp.up_proj.norm.weight` | `blk.{i}.ffn_up_norm_in.weight` | BitNet per-projection | +| `layers.{i}.mlp.down_proj.norm.weight` | `blk.{i}.ffn_down_norm_in.weight` | BitNet per-projection | + +### Architecture-Specific Tensors + +**Qwen3:** + +| HF Name | GGUF Name | +|----------|-----------| +| `layers.{i}.post_attention_layernorm.weight` | `blk.{i}.ffn_norm.weight` | + +**Gemma3 (additional):** + +| HF Name | GGUF Name | +|----------|-----------| +| `layers.{i}.post_attention_layernorm.weight` | `blk.{i}.post_attention_norm.weight` | +| `layers.{i}.pre_feedforward_layernorm.weight` | `blk.{i}.ffn_norm.weight` | +| `layers.{i}.post_feedforward_layernorm.weight` | `blk.{i}.post_ffw_norm.weight` | + +--- + +## 3. Conversion Script + +### `utils/convert-bitnet-embedding-to-gguf.py` + +Unified standalone conversion script (safetensors → GGUF) that **auto-detects** the model architecture from `config.json`'s `model_type` field (`qwen3` or `gemma3_text`). Key features: + +- Hardcoded HF→GGUF tensor name mapping (no dependency on llama.cpp's Python converter) +- Auto-detection of architecture and GGUF arch string (`qwen3` / `gemma3`) +- Supports three output types: + - `--outtype f32`: all weights in float32 + - `--outtype f16`: 2D weights and embeddings as float16, norms as float16 + - `--outtype i2_s`: ternary weights packed in I2_S layout, non-ternary weights as float16 +- Writes `key_length` and `value_length` metadata for correct head_dim (critical: head_dim != hidden_size/num_heads for both models, default calculation would give wrong values) +- BPE tokenizer handling with per-architecture pre-tokenizer hash verification: + - Qwen3: GPT-2 BPE tokenizer + - Gemma3: GemmaTokenizerFast (BPE) +- Pooling type auto-detection from `modules.json` / `1_Pooling/config.json` (sentence-transformers convention) +- Architecture-specific tokenizer handling: + - Qwen3: EOS token override (`<|endoftext|>` 151643) + `add_eos_token(True)` for last-token pooling + - Gemma3: EOS token auto-set by SpecialVocab from tokenizer_config.json (eos_token_id=1) +- Gemma3: writes `query_pre_attn_scalar = 256` for correct attention scaling + +### I2_S Ternary Packing + +The I2_S format packs ternary weights {-1, 0, +1} into 2-bit representation: + +- Quantization: `scale = 1/mean(|w|)`, `q = round(w * scale).clamp(-1, 1)` +- Encoding: `-1 → 0`, `0 → 1`, `+1 → 2` +- Every 128 values form a block, packed into 32 bytes +- Each byte stores 4 values: `byte = (c0 << 6) | (c1 << 4) | (c2 << 2) | c3` +- Scale (float32) is appended at the end of the packed data buffer + +### Tensor Type Assignment + +| Tensor Type | f16 mode | i2_s mode | +|-------------|----------|-----------| +| 2D linear weights | float16 | I2_S ternary packed | +| Embedding weights | float16 | float16 | +| Norm weights (1D) | float16 | float16 | + +Note: `output.weight` (lm_head) is skipped for embedding models — it is not needed (no token generation). + +--- + +## 4. C++ Modifications (`3rdparty/llama.cpp/src/llama.cpp`) + +### 4.1 New Architecture: `LLM_ARCH_GEMMA3` + +Added after `LLM_ARCH_GEMMA2` in the `llm_arch` enum with name mapping `"gemma3"`. Qwen3 (`LLM_ARCH_QWEN3`) was added by the 0.6b adaptation. + +### 4.2 New Tensor Enums (shared across architectures) + +Added 7 new entries after `LLM_TENSOR_FFN_SUB_NORM`: + +```cpp +LLM_TENSOR_ATTN_Q_NORM_IN, +LLM_TENSOR_ATTN_K_NORM_IN, +LLM_TENSOR_ATTN_V_NORM_IN, +LLM_TENSOR_ATTN_OUT_NORM_IN, +LLM_TENSOR_FFN_GATE_NORM_IN, +LLM_TENSOR_FFN_UP_NORM_IN, +LLM_TENSOR_FFN_DOWN_NORM_IN, +``` + +### 4.3 Layer Struct Fields + +Added to `struct llama_layer`: + +```cpp +struct ggml_tensor * attn_q_norm_in; +struct ggml_tensor * attn_k_norm_in; +struct ggml_tensor * attn_v_norm_in; +struct ggml_tensor * attn_out_norm_in; +struct ggml_tensor * ffn_gate_norm_in; +struct ggml_tensor * ffn_up_norm_in; +struct ggml_tensor * ffn_down_norm_in; +``` + +### 4.4 Tensor Name Mappings + +Both `LLM_ARCH_QWEN3` and `LLM_ARCH_GEMMA3` include the 7 per-projection norm tensor mappings plus standard tensors (see Section 2 for full mapping). Key differences: + +- Qwen3 includes `LLM_TENSOR_OUTPUT` (`"output"`); Gemma3 does not (uses tied embeddings directly) +- Gemma3 additionally includes `LLM_TENSOR_ATTN_POST_NORM` (`"blk.%d.post_attention_norm"`) and `LLM_TENSOR_FFN_POST_NORM` (`"blk.%d.post_ffw_norm"`) + +### 4.5 load_tensors + +Both architectures load the 7 per-projection norm tensors as optional (`TENSOR_NOT_REQUIRED`): + +```cpp +layer.attn_q_norm_in = create_tensor(tn(...), {n_embd}, TENSOR_NOT_REQUIRED); +layer.attn_k_norm_in = create_tensor(tn(...), {n_embd}, TENSOR_NOT_REQUIRED); +layer.attn_v_norm_in = create_tensor(tn(...), {n_embd}, TENSOR_NOT_REQUIRED); +layer.attn_out_norm_in = create_tensor(tn(...), {n_embd_head_k * n_head}, TENSOR_NOT_REQUIRED); +layer.ffn_gate_norm_in = create_tensor(tn(...), {n_embd}, TENSOR_NOT_REQUIRED); +layer.ffn_up_norm_in = create_tensor(tn(...), {n_embd}, TENSOR_NOT_REQUIRED); +layer.ffn_down_norm_in = create_tensor(tn(...), {n_ff}, TENSOR_NOT_REQUIRED); +``` + +Note: `o_proj.norm` input dimension is `n_embd_head_k * n_head` (Qwen3: 2048, Gemma3: 1024), `down_proj.norm` input dimension is `n_ff` (Qwen3: 3072, Gemma3: 2048). + +Both graph functions use the same per-projection norm pattern. The logic is fully backward compatible — when no `*_norm_in` tensors exist, behavior is identical to the original. + +**Attention per-projection norms:** +``` +// Before Q/K/V matmul: +if (layer.attn_q_norm_in) { + cur_q = ggml_rms_norm(ctx, cur, hparams.f_norm_rms_eps); + cur_q = ggml_mul(ctx, cur_q, layer.attn_q_norm_in); +} else { + cur_q = cur; +} +Qcur = ggml_mul_mat(ctx, layer.wq, cur_q); +// QK head norms applied after projection +Qcur = ggml_rms_norm(ctx, Qcur, hparams.f_norm_rms_eps); +Qcur = ggml_mul(ctx, Qcur, layer.attn_q_norm); +``` + +**O_proj norm** requires special handling because `llm_build_kv()` normally applies `wo` internally. Solution: pass `wo=NULL` to `llm_build_kv()`, then apply norm + wo manually: + +``` +cur = llm_build_kv(..., wo=NULL, ...); // returns attention output without o_proj +if (layer.attn_out_norm_in) { + cur = ggml_rms_norm(ctx, cur, hparams.f_norm_rms_eps); + cur = ggml_mul(ctx, cur, layer.attn_out_norm_in); +} +cur = ggml_mul_mat(ctx, layer.wo, cur); +``` + +**FFN per-projection norms:** +``` +// Instead of llm_build_ffn(), manually: +if (layer.ffn_gate_norm_in) { + tmp_gate = rms_norm(cur) * gate_norm_in; +} else { + tmp_gate = cur; +} +tmp_gate = matmul(gate_proj, tmp_gate); +tmp_gate = activation(tmp_gate); // SiLU for Qwen3, GELU for Gemma3 +// Similarly for up_proj +tmp = tmp_gate * tmp_up; + +if (layer.ffn_down_norm_in) { + tmp = rms_norm(tmp) * down_norm_in; +} +cur = matmul(down_proj, tmp); +``` + +**Gemma3-specific differences:** +- Embedding scaling by `sqrt(n_embd)` (Gemma convention) +- GELU activation instead of SiLU +- Post-attention and post-FFN layer norms +- `query_pre_attn_scalar` for attention scaling + +--- + +## 5. GGUF Conversion Process + +Each model variant requires two GGUF files from **two different source models**: + +### 5.1 Qwen3 (0.6b) + +| GGUF Output | Source Model | Description | +|-------------|-------------|-------------| +| `embeddings-0.6b-f16.gguf` | `multilingual-e5-0.6b` (standard Qwen3) | F16 baseline | +| `bitnet-embeddings-0.6b-f16-i2_s.gguf` | `bitnet-embeddings-0.6b` (BitNet ternary) | I2_S ternary packed | + +**F16 (from standard Qwen3 model):** +```bash +python3 utils/convert-bitnet-embedding-to-gguf.py \ + /path/to/multilingual-e5-0.6b \ + --outtype f16 \ + --outfile embeddings-0.6b-f16.gguf +``` + +What happens: +1. Load `model.safetensors` (standard Qwen3 weights, bfloat16) +2. Convert all 2D weights (projections, embeddings) to float16 +3. Convert norm weights to float16 +4. Write GGUF with `qwen3` architecture metadata and tokenizer + +**Output:** ~1.11 GiB (595.78M params) + +**I2_S (from BitNet model):** +```bash +python3 utils/convert-bitnet-embedding-to-gguf.py \ + /path/to/bitnet-embeddings-0.6b \ + --outfile bitnet-embeddings-0.6b-f16-i2_s.gguf --outtype i2_s +``` + +What happens: +1. Load `model.safetensors` (BitNet ternary weights, bfloat16) +2. Map HF tensor names to GGUF names, including 7 extra `*_norm_in` tensors per layer +3. For each 2D linear weight: quantize to I2_S ternary packed format +4. Keep embeddings (`token_embd.weight`) in float16 +5. Keep all norm weights in float16 +6. Skip `output.weight` (lm_head, not needed for embedding models) +7. Write GGUF with `I2_S` type tag for quantized tensors + +**Output:** ~699 MiB (~50% of F16 size) + +### 5.2 Gemma3 (270m) + +| GGUF Output | Source Model | Description | +|-------------|-------------|-------------| +| `multilingual-e5-270m-f16.gguf` | `multilingual-e5-270m-260311` (standard Gemma3) | F16 baseline | +| `bitnet-embeddings-270m-i2_s.gguf` | `bitnet-embeddings-270m` (BitNet ternary) | I2_S ternary packed | + +**F16 (from standard Gemma3 model):** +```bash +python3 utils/convert-bitnet-embedding-to-gguf.py \ + /path/to/multilingual-e5-270m-260311 \ + --outtype f16 +``` + +What happens: +1. Load `model.safetensors` (standard Gemma3 weights, bfloat16) +2. Convert all 2D weights (projections, embeddings) to float16 +3. Convert norm weights to float16 +4. Write GGUF with `gemma3` architecture metadata and tokenizer + +**I2_S (from BitNet model):** +```bash +python3 utils/convert-bitnet-embedding-to-gguf.py \ + /path/to/bitnet-embeddings-270m \ + --outtype i2_s +``` + +What happens: +1. Load `model.safetensors` (BitNet ternary weights, bfloat16) +2. Map HF tensor names to GGUF names, including 7 extra `*_norm_in` tensors per layer +3. For each 2D linear weight: quantize to I2_S ternary packed format +4. Keep embeddings (`token_embd.weight`) in float16 +5. Keep all norm weights in float16 +6. Skip `output.weight` (lm_head, not needed for embedding models) +7. Write GGUF with `I2_S` type tag for quantized tensors + +### 5.3 Why Two Different Source Models? + +- `multilingual-e5-*` is the **teacher/baseline model** with standard float weights, used as the F16 performance reference +- `bitnet-embeddings-*` is the **1-bit quantized student model** with ternary weights and per-projection BitLinear norms, converted to I2_S for efficient CPU inference +- Benchmarking compares both to measure the throughput gain and quality trade-off of ternary quantization + +### 5.4 Tensor Type Summary + +| Tensor | F16 (baseline) | I2_S (BitNet) | +|--------|----------------|---------------| +| Linear projections (q/k/v/o/gate/up/down) | float16 | I2_S (2-bit packed + float32 scale) | +| Embedding (`token_embd.weight`) | float16 | float16 | +| Per-projection norms (`*_norm_in`) | N/A (not present) | float16 | +| Layer norms (attn_norm, ffn_norm, etc.) | float16 | float16 | +| QK head norms (`attn_q_norm`, `attn_k_norm`) | float16 | float16 | +| `output.weight` (lm_head) | skipped | skipped | + +--- + +## 6. Additional Changes + +### 6.1 ggml.c: F16 Norm Weight Support + +Added `ggml_compute_forward_mul_f32_f16()` function to support element-wise multiplication where norm weights are stored in float16. Modified `ggml_compute_forward_mul()` to dispatch based on `src1->type`. + +### 6.2 gguf-py: I2_S Type + +Added `I2_S = 36` to `GGMLQuantizationType` enum and `(4, 1)` quant size in `constants.py`. + +### 6.3 CMakeLists.txt: BitNet LUT Kernels Guard + +Guarded `bitnet-lut-kernels.h` include with `if (GGML_BITNET_ARM_TL1 OR GGML_BITNET_X86_TL2)` to prevent build errors when LUT kernels are not available. + +### 6.4 ggml-bitnet-mad.cpp: AVX512 SIMD + +Added AVX512BW SIMD paths for I2_S dot product functions: +- `ggml_vec_dot_i2_i8_s_1x1` +- `ggml_vec_dot_i2_i8_s_1xN` +- `ggml_vec_dot_i2_i8_s_Nx1` + +--- + +## 7. Build and Run + +```bash +# Build with BitNet repo (includes I2_S support) +cmake -S /path/to/BitNet -B build -DCMAKE_BUILD_TYPE=Release +cmake --build build --target llama-embedding llama-bench -j$(nproc) + +# Run embedding inference (Qwen3 example) +build/bin/llama-embedding -m bitnet-embeddings-0.6b-f16-i2_s.gguf \ + -p "hello world" --embd-normalize 2 --embd-output-format array + +# Run embedding inference (Gemma3 example) +build/bin/llama-embedding -m bitnet-embeddings-270m-i2_s.gguf \ + -p "hello world" --embd-normalize 2 --embd-output-format array + +# Benchmark: F16 vs I2_S (Qwen3) +build/bin/llama-bench -m embeddings-0.6b-f16.gguf \ + -t 8 -p 128,256,512,1024,2048 -n 32,64 -r 3 -ngl 0 + +build/bin/llama-bench -m bitnet-embeddings-0.6b-f16-i2_s.gguf \ + -t 8 -p 128,256,512,1024,2048 -n 32,64 -r 3 -ngl 0 + +# Benchmark: F16 vs I2_S (Gemma3) +build/bin/llama-bench -m multilingual-e5-270m-f16.gguf \ + -t 8 -p 128,256,512,1024,2048 -n 32,64 -r 3 -ngl 0 + +build/bin/llama-bench -m bitnet-embeddings-270m-i2_s.gguf \ + -t 8 -p 128,256,512,1024,2048 -n 32,64 -r 3 -ngl 0 +``` diff --git a/docs/bitnet-embeddings-qwen3-gguf-conversion.md b/docs/bitnet-embeddings-qwen3-gguf-conversion.md deleted file mode 100644 index 9d63c93..0000000 --- a/docs/bitnet-embeddings-qwen3-gguf-conversion.md +++ /dev/null @@ -1,302 +0,0 @@ -# BitNet Embeddings (Qwen3) GGUF Conversion Implementation - -## 1. Background - -`bitnet-embeddings-0.6b` is a Qwen3-based embedding model with BitNet per-projection RMSNorm (`BitLinear`). Each linear projection (q/k/v/o/gate/up/down) has a `.norm.weight` that applies RMSNorm to the input **before** the matmul: - -``` -x → RMSNorm(x, norm.weight) → activation_quant(8bit) → matmul(weight_quant(ternary)) -``` - -This pattern does **not** exist in any standard llama.cpp architecture: -- Standard Qwen3: no per-projection norms -- Standard BitNet: has `attn_sub_norm`/`ffn_sub_norm` at different positions (after attention/gate*up, not before each projection) - -### Model Config - -- Architecture: `Qwen3Model` -- hidden_size: 1024, num_attention_heads: 16, num_key_value_heads: 8 -- head_dim: 128 (note: != hidden_size/num_heads = 64) -- intermediate_size: 3072, num_hidden_layers: 28 -- tie_word_embeddings: true -- rope_theta: 1000000, rms_norm_eps: 1e-06 - -### Per-Layer Tensors (7 extra norm tensors per layer) - -| Tensor | Shape | -|--------|-------| -| `self_attn.q_proj.norm.weight` | [1024] | -| `self_attn.k_proj.norm.weight` | [1024] | -| `self_attn.v_proj.norm.weight` | [1024] | -| `self_attn.o_proj.norm.weight` | [2048] | -| `mlp.gate_proj.norm.weight` | [1024] | -| `mlp.up_proj.norm.weight` | [1024] | -| `mlp.down_proj.norm.weight` | [3072] | - ---- - -## 2. GGUF Tensor Name Mapping - -| HF Name | GGUF Name | Notes | -|----------|-----------|-------| -| `embed_tokens.weight` | `token_embd.weight` | | -| `norm.weight` | `output_norm.weight` | | -| `layers.{i}.input_layernorm.weight` | `blk.{i}.attn_norm.weight` | | -| `layers.{i}.post_attention_layernorm.weight` | `blk.{i}.ffn_norm.weight` | | -| `layers.{i}.self_attn.q_proj.weight` | `blk.{i}.attn_q.weight` | | -| `layers.{i}.self_attn.k_proj.weight` | `blk.{i}.attn_k.weight` | | -| `layers.{i}.self_attn.v_proj.weight` | `blk.{i}.attn_v.weight` | | -| `layers.{i}.self_attn.o_proj.weight` | `blk.{i}.attn_output.weight` | | -| `layers.{i}.self_attn.q_norm.weight` | `blk.{i}.attn_q_norm.weight` | QK head norm | -| `layers.{i}.self_attn.k_norm.weight` | `blk.{i}.attn_k_norm.weight` | QK head norm | -| `layers.{i}.self_attn.q_proj.norm.weight` | `blk.{i}.attn_q_norm_in.weight` | BitNet per-projection | -| `layers.{i}.self_attn.k_proj.norm.weight` | `blk.{i}.attn_k_norm_in.weight` | BitNet per-projection | -| `layers.{i}.self_attn.v_proj.norm.weight` | `blk.{i}.attn_v_norm_in.weight` | BitNet per-projection | -| `layers.{i}.self_attn.o_proj.norm.weight` | `blk.{i}.attn_output_norm_in.weight` | BitNet per-projection | -| `layers.{i}.mlp.gate_proj.weight` | `blk.{i}.ffn_gate.weight` | | -| `layers.{i}.mlp.up_proj.weight` | `blk.{i}.ffn_up.weight` | | -| `layers.{i}.mlp.down_proj.weight` | `blk.{i}.ffn_down.weight` | | -| `layers.{i}.mlp.gate_proj.norm.weight` | `blk.{i}.ffn_gate_norm_in.weight` | BitNet per-projection | -| `layers.{i}.mlp.up_proj.norm.weight` | `blk.{i}.ffn_up_norm_in.weight` | BitNet per-projection | -| `layers.{i}.mlp.down_proj.norm.weight` | `blk.{i}.ffn_down_norm_in.weight` | BitNet per-projection | - ---- - -## 3. Conversion Script - -### `utils/convert-bitnet-embedding-to-gguf.py` - -Standalone conversion script (safetensors → GGUF). Key features: - -- Hardcoded HF→GGUF tensor name mapping (no dependency on llama.cpp's Python converter) -- Supports three output types: - - `--outtype f32`: all weights in float32 - - `--outtype f16`: 2D weights and embeddings as float16, norms as float16 - - `--outtype i2_s`: ternary weights packed in I2_S layout, non-ternary weights as float16 -- Writes `key_length` and `value_length` metadata for head_dim=128 (critical: default calculation would give wrong value 64) -- GPT-2 BPE tokenizer handling with pre-tokenizer hash verification -- Pooling type auto-detection from `modules.json` / `1_Pooling/config.json` (sentence-transformers convention) -- EOS token override: uses `<|endoftext|>` (151643) for correct last-token pooling -- Architecture string: `"qwen3"` - -### I2_S Ternary Packing - -The I2_S format packs ternary weights {-1, 0, +1} into 2-bit representation: - -- Quantization: `scale = 1/mean(|w|)`, `q = round(w * scale).clamp(-1, 1)` -- Encoding: `-1 → 0`, `0 → 1`, `+1 → 2` -- Every 128 values form a block, packed into 32 bytes -- Each byte stores 4 values: `byte = (c0 << 6) | (c1 << 4) | (c2 << 2) | c3` -- Scale (float32) is appended at the end of the packed data buffer - -### Tensor Type Assignment - -| Tensor Type | f16 mode | i2_s mode | -|-------------|----------|-----------| -| 2D linear weights | float16 | I2_S ternary packed | -| Embedding weights | float16 | float16 | -| Norm weights (1D) | float16 | float16 | - -Note: `output.weight` (lm_head) is skipped for embedding models — it is not needed (no token generation). - ---- - -## 4. C++ Modifications (`3rdparty/llama.cpp/src/llama.cpp`) - -### 4.1 New Tensor Enums - -Added 7 new entries after `LLM_TENSOR_FFN_SUB_NORM`: - -```cpp -LLM_TENSOR_ATTN_Q_NORM_IN, -LLM_TENSOR_ATTN_K_NORM_IN, -LLM_TENSOR_ATTN_V_NORM_IN, -LLM_TENSOR_ATTN_OUT_NORM_IN, -LLM_TENSOR_FFN_GATE_NORM_IN, -LLM_TENSOR_FFN_UP_NORM_IN, -LLM_TENSOR_FFN_DOWN_NORM_IN, -``` - -### 4.2 Tensor Name Mappings - -Added to `LLM_ARCH_QWEN3` tensor name map: - -```cpp -{ LLM_TENSOR_ATTN_Q_NORM_IN, "blk.%d.attn_q_norm_in" }, -{ LLM_TENSOR_ATTN_K_NORM_IN, "blk.%d.attn_k_norm_in" }, -{ LLM_TENSOR_ATTN_V_NORM_IN, "blk.%d.attn_v_norm_in" }, -{ LLM_TENSOR_ATTN_OUT_NORM_IN, "blk.%d.attn_output_norm_in" }, -{ LLM_TENSOR_FFN_GATE_NORM_IN, "blk.%d.ffn_gate_norm_in" }, -{ LLM_TENSOR_FFN_UP_NORM_IN, "blk.%d.ffn_up_norm_in" }, -{ LLM_TENSOR_FFN_DOWN_NORM_IN, "blk.%d.ffn_down_norm_in" }, -``` - -### 4.3 Layer Struct Fields - -Added to `struct llama_layer`: - -```cpp -struct ggml_tensor * attn_q_norm_in; -struct ggml_tensor * attn_k_norm_in; -struct ggml_tensor * attn_v_norm_in; -struct ggml_tensor * attn_out_norm_in; -struct ggml_tensor * ffn_gate_norm_in; -struct ggml_tensor * ffn_up_norm_in; -struct ggml_tensor * ffn_down_norm_in; -``` - -### 4.4 load_tensors (LLM_ARCH_QWEN3) - -Added optional loading with `TENSOR_NOT_REQUIRED`: - -```cpp -layer.attn_q_norm_in = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM_IN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED); -layer.attn_k_norm_in = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM_IN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED); -layer.attn_v_norm_in = create_tensor(tn(LLM_TENSOR_ATTN_V_NORM_IN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED); -layer.attn_out_norm_in = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM_IN, "weight", i), {n_embd_head_k * n_head}, TENSOR_NOT_REQUIRED); -layer.ffn_gate_norm_in = create_tensor(tn(LLM_TENSOR_FFN_GATE_NORM_IN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED); -layer.ffn_up_norm_in = create_tensor(tn(LLM_TENSOR_FFN_UP_NORM_IN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED); -layer.ffn_down_norm_in = create_tensor(tn(LLM_TENSOR_FFN_DOWN_NORM_IN, "weight", i), {n_ff}, TENSOR_NOT_REQUIRED); -``` - -Note: `o_proj.norm` input dimension is `n_embd_head_k * n_head` (=2048), `down_proj.norm` input dimension is `n_ff` (=3072). - -### 4.5 build_qwen3() Graph Modifications - -The `build_qwen3()` function was modified to conditionally apply per-projection RMSNorm. The logic is fully backward compatible — when no `*_norm_in` tensors exist, behavior is identical to original. - -**Attention per-projection norms:** -``` -// Before Q/K/V matmul: -if (layer.attn_q_norm_in) { - cur_q = ggml_rms_norm(ctx, cur, hparams.f_norm_rms_eps); - cur_q = ggml_mul(ctx, cur_q, layer.attn_q_norm_in); -} else { - cur_q = cur; -} -Qcur = ggml_mul_mat(ctx, layer.wq, cur_q); -// Similarly for K, V -``` - -**O_proj norm** requires special handling because `llm_build_kv()` normally applies `wo` internally. Solution: pass `wo=NULL` to `llm_build_kv()`, then apply norm + wo manually: - -``` -cur = llm_build_kv(..., wo=NULL, ...); // returns attention output without o_proj -if (layer.attn_out_norm_in) { - cur = ggml_rms_norm(ctx, cur, hparams.f_norm_rms_eps); - cur = ggml_mul(ctx, cur, layer.attn_out_norm_in); -} -cur = ggml_mul_mat(ctx, layer.wo, cur); -``` - -**FFN per-projection norms:** -``` -// Instead of llm_build_ffn(), manually: -if (layer.ffn_gate_norm_in) { - tmp_gate = rms_norm(cur) * gate_norm_in; -} else { - tmp_gate = cur; -} -tmp_gate = matmul(gate_proj, tmp_gate); -// Similarly for up_proj -tmp = silu(tmp_gate) * tmp_up; - -if (layer.ffn_down_norm_in) { - tmp = rms_norm(tmp) * down_norm_in; -} -cur = matmul(down_proj, tmp); -``` - ---- - -## 5. GGUF Conversion Process - -There are two GGUF files to produce, from **two different source models**: - -| GGUF Output | Source Model | Description | -|-------------|-------------|-------------| -| `embeddings-0.6b-f16.gguf` | `multilingual-e5-0.6b` (standard Qwen3) | F16 baseline, standard float16 weights | -| `bitnet-embeddings-0.6b-f16-i2_s.gguf` | `bitnet-embeddings-0.6b` (BitNet ternary) | I2_S ternary packed weights | - -### 5.1 F16 GGUF: from multilingual-e5-0.6b - -The F16 GGUF is converted from the **standard (non-BitNet) model** `multilingual-e5-0.6b`, which has normal float weights and no per-projection RMSNorm. This uses llama.cpp's standard converter since it is a vanilla Qwen3 model: - -```bash -python3 /path/to/llama.cpp/convert_hf_to_gguf.py \ - /path/to/multilingual-e5-0.6b \ - --outtype f16 \ - --outfile embeddings-0.6b-f16.gguf -``` - -**What happens:** -1. Load `model.safetensors` (standard Qwen3 weights, bfloat16) -2. Convert all 2D weights (projections, embeddings) to float16 -3. Convert norm weights to float32 -4. Write GGUF with `qwen3` architecture metadata and tokenizer - -**Output:** ~1.11 GiB (595.78M params) - -### 5.2 I2_S GGUF: from bitnet-embeddings-0.6b - -The I2_S GGUF is converted from the **BitNet ternary model** `bitnet-embeddings-0.6b`, which has ternary weights {-1, 0, +1} and 7 extra per-projection RMSNorm tensors per layer. This uses the custom converter because the standard llama.cpp converter does not handle per-projection norms or I2_S quantization: - -```bash -python3 utils/convert-bitnet-embedding-to-gguf.py \ - /path/to/bitnet-embeddings-0.6b \ - --outfile bitnet-embeddings-0.6b-f16-i2_s.gguf --outtype i2_s -``` - -**What happens:** -1. Load `model.safetensors` (BitNet ternary weights, bfloat16) -2. Map HF tensor names to GGUF names, including 7 extra `*_norm_in` tensors per layer (see Section 2) -3. For each 2D linear weight (q/k/v/o/gate/up/down projections): - - Compute scale: `scale = 1 / mean(|w|)` - - Quantize: `q = round(w * scale).clamp(-1, 1)` - - Encode: `-1 -> 0`, `0 -> 1`, `+1 -> 2` - - Pack every 128 values into 32 bytes (4 values per byte, 2 bits each) - - Append per-row float32 scale -4. Keep embeddings (`token_embd.weight`) in float16 (not ternary) -5. Keep all norm weights in float16 -6. Skip `output.weight` (lm_head, not needed for embedding models) -7. Write GGUF with `I2_S` type tag for quantized tensors - -**Output:** ~699 MiB (~50% of F16 size) - -### 5.3 Why Two Different Source Models? - -- `multilingual-e5-0.6b` is the **teacher/baseline model** with standard float weights, used as the F16 performance reference -- `bitnet-embeddings-0.6b` is the **1-bit quantized student model** with ternary weights and per-projection BitLinear norms, converted to I2_S for efficient CPU inference -- Benchmarking compares both to measure the throughput gain and quality trade-off of ternary quantization - -### 5.4 Tensor Type Summary - -| Tensor | F16 (from e5-0.6b) | I2_S (from bitnet-0.6b) | -|--------|---------------------|-------------------------| -| Linear projections (q/k/v/o/gate/up/down) | float16 | I2_S (2-bit packed + float32 scale) | -| Embedding (`token_embd.weight`) | float16 | float16 | -| Per-projection norms (`*_norm_in`) | N/A (not present) | float16 | -| Layer norms (`attn_norm`, `ffn_norm`) | float32 | float16 | -| QK head norms (`attn_q_norm`, `attn_k_norm`) | float32 | float32 | -| `output.weight` (lm_head) | present | skipped | - ---- - -## 6. Build and Run - -```bash -# Build with BitNet repo (includes I2_S support) -cmake -S /path/to/BitNet -B build -DCMAKE_BUILD_TYPE=Release -cmake --build build --target llama-embedding llama-bench -j$(nproc) - -# Run embedding inference -build/bin/llama-embedding -m bitnet-embeddings-0.6b-f16-i2_s.gguf \ - -p "hello world" --embd-normalize 2 --embd-output-format array - -# Benchmark: F16 vs I2_S -build/bin/llama-bench -m embeddings-0.6b-f16.gguf \ - -t 8 -p 128,256,512,1024,2048 -n 32,64 -r 3 -ngl 0 - -build/bin/llama-bench -m bitnet-embeddings-0.6b-f16-i2_s.gguf \ - -t 8 -p 128,256,512,1024,2048 -n 32,64 -r 3 -ngl 0 -``` diff --git a/utils/convert-bitnet-embedding-to-gguf.py b/utils/convert-bitnet-embedding-to-gguf.py index 3a43407..3a9e489 100644 --- a/utils/convert-bitnet-embedding-to-gguf.py +++ b/utils/convert-bitnet-embedding-to-gguf.py @@ -23,11 +23,17 @@ import gguf logger = logging.getLogger("convert-bitnet-embedding") +# Supported architectures: model_type -> gguf arch name +SUPPORTED_ARCHS = { + "qwen3": "qwen3", + "gemma3_text": "gemma3", +} + # --------------------------------------------------------------------------- # Tensor name mapping: HuggingFace -> GGUF # --------------------------------------------------------------------------- -def build_tensor_name_map(n_layers: int) -> dict[str, str]: +def build_tensor_name_map(n_layers: int, arch: str) -> dict[str, str]: """Build HF tensor name -> GGUF tensor name mapping.""" mapping: dict[str, str] = { "embed_tokens.weight": "token_embd.weight", @@ -41,7 +47,6 @@ def build_tensor_name_map(n_layers: int) -> dict[str, str]: mapping.update({ # Layer norms f"{pfx}.input_layernorm.weight": f"{blk}.attn_norm.weight", - f"{pfx}.post_attention_layernorm.weight": f"{blk}.ffn_norm.weight", # Self-attention projections f"{pfx}.self_attn.q_proj.weight": f"{blk}.attn_q.weight", @@ -49,7 +54,7 @@ def build_tensor_name_map(n_layers: int) -> dict[str, str]: f"{pfx}.self_attn.v_proj.weight": f"{blk}.attn_v.weight", f"{pfx}.self_attn.o_proj.weight": f"{blk}.attn_output.weight", - # QK head norms (standard Qwen3) + # QK head norms f"{pfx}.self_attn.q_norm.weight": f"{blk}.attn_q_norm.weight", f"{pfx}.self_attn.k_norm.weight": f"{blk}.attn_k_norm.weight", @@ -70,20 +75,29 @@ def build_tensor_name_map(n_layers: int) -> dict[str, str]: f"{pfx}.mlp.down_proj.norm.weight": f"{blk}.ffn_down_norm_in.weight", }) + if arch == "qwen3": + mapping[f"{pfx}.post_attention_layernorm.weight"] = f"{blk}.ffn_norm.weight" + elif arch == "gemma3_text": + mapping.update({ + f"{pfx}.post_attention_layernorm.weight": f"{blk}.post_attention_norm.weight", + f"{pfx}.pre_feedforward_layernorm.weight": f"{blk}.ffn_norm.weight", + f"{pfx}.post_feedforward_layernorm.weight": f"{blk}.post_ffw_norm.weight", + }) + return mapping # --------------------------------------------------------------------------- -# Tokenizer handling (GPT-2 / BPE for Qwen3) +# Tokenizer handling # --------------------------------------------------------------------------- -def get_vocab_base_pre(tokenizer) -> str: +def get_vocab_base_pre(tokenizer, arch: str) -> str: # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that # is specific for the BPE pre-tokenizer used by the model # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can # use in llama.cpp to implement the same pre-tokenizer - chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n\U0001f680 (normal) \U0001f636‍\U0001f32b️ (multiple emojis concatenated) ✅ \U0001f999\U0001f999 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច\U0001f601 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````""""......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL' + chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n\U0001f680 (normal) \U0001f636‍\U0001f32b️ (multiple emojis concatenated) ✅ \U0001f999\U0001f999 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច\U0001f601 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````""""""......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL' chktok = tokenizer.encode(chktxt) chkhsh = sha256(str(chktok).encode()).hexdigest() @@ -93,27 +107,38 @@ def get_vocab_base_pre(tokenizer) -> str: res = None - # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script - # or pull the latest version of the model from Huggingface - # don't edit the hashes manually! - if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5": - # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B - res = "llama-bpe" - if chkhsh == "049ecf7629871e3041641907f3de7c733e4dbfdc736f57d882ba0b0845599754": - # ref: https://huggingface.co/deepseek-ai/deepseek-llm-7b-base - res = "deepseek-llm" - if chkhsh == "347715f544604f9118bb75ed199f68779f423cabb20db6de6f31b908d04d7821": - # ref: https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base - res = "deepseek-coder" - if chkhsh == "8aeee3860c56296a157a1fe2fad249ec40aa59b1bb5709f4ade11c4e6fe652ed": - # ref: https://huggingface.co/tiiuae/falcon-7b - res = "falcon" - if chkhsh == "3ce83efda5659b07b1ad37ca97ca5797ea4285d9b9ab0dc679e4a720c9da7454": - # ref: https://huggingface.co/openai-community/gpt2 - res = "gpt-2" - if chkhsh == "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c": - # ref: https://huggingface.co/Qwen/Qwen3-Embedding-0.6B - res = "qwen2" + if arch == "qwen3": + # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script + # or pull the latest version of the model from Huggingface + # don't edit the hashes manually! + if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5": + # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B + res = "llama-bpe" + if chkhsh == "049ecf7629871e3041641907f3de7c733e4dbfdc736f57d882ba0b0845599754": + # ref: https://huggingface.co/deepseek-ai/deepseek-llm-7b-base + res = "deepseek-llm" + if chkhsh == "347715f544604f9118bb75ed199f68779f423cabb20db6de6f31b908d04d7821": + # ref: https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base + res = "deepseek-coder" + if chkhsh == "8aeee3860c56296a157a1fe2fad249ec40aa59b1bb5709f4ade11c4e6fe652ed": + # ref: https://huggingface.co/tiiuae/falcon-7b + res = "falcon" + if chkhsh == "3ce83efda5659b07b1ad37ca97ca5797ea4285d9b9ab0dc679e4a720c9da7454": + # ref: https://huggingface.co/openai-community/gpt2 + res = "gpt-2" + if chkhsh == "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c": + # ref: https://huggingface.co/Qwen/Qwen3-Embedding-0.6B + res = "qwen2" + if chkhsh == "855d9fb74bb0b28ce2305e9cd037ff6d8c798f18d19381ddfc14bea3dc9c002f": + # ref: multilingual-e5-0.6b-260311 (Qwen3 tokenizer variant) + res = "qwen2" + elif arch == "gemma3_text": + if chkhsh == "fcb6bf9f20f6c40fa4aa4f7f99607bd6c106ca2348efdacacdca8152e59dcfe9": + # ref: multilingual-e5-270m-260311 (Gemma3 tokenizer) + res = "default" + if chkhsh == "a8594e3edff7c29c003940395316294b2c623571571fc8d3d2d6571f5571cbe6": + # ref: google/gemma-2-9b + res = "default" if res is None: logger.warning("\n") @@ -146,13 +171,29 @@ def _does_token_look_special(token: str) -> bool: return False -def set_vocab(gguf_writer: gguf.GGUFWriter, dir_model: Path, hparams: dict): - """Set GPT-2 BPE vocab for Qwen3.""" +def set_vocab(gguf_writer: gguf.GGUFWriter, dir_model: Path, hparams: dict, arch: str): + """Set tokenizer vocab. + + - Qwen3: BPE tokenizer (tokenizer.ggml.model = "gpt2") + - Gemma3: SPM-compatible tokenizer from tokenizer.json (tokenizer.ggml.model = "llama") + Gemma uses SentencePiece-style tokenization with ▁ space prefix and byte fallback. + Using "llama" model type ensures llama.cpp uses the correct SPM pre-tokenizer + instead of the BPE regex-based pre-tokenizer which breaks CJK tokenization. + """ from transformers import AutoTokenizer tokenizer = AutoTokenizer.from_pretrained(dir_model) vocab_size = hparams.get("vocab_size", len(tokenizer.vocab)) - tokpre = get_vocab_base_pre(tokenizer) + if arch == "gemma3_text": + _set_vocab_gemma3(gguf_writer, dir_model, tokenizer, vocab_size) + else: + _set_vocab_bpe(gguf_writer, dir_model, tokenizer, vocab_size, arch) + + +def _set_vocab_bpe(gguf_writer: gguf.GGUFWriter, dir_model: Path, + tokenizer, vocab_size: int, arch: str): + """Set BPE vocab (for Qwen3).""" + tokpre = get_vocab_base_pre(tokenizer, arch) tokens: list[str] = [] toktypes: list[int] = [] @@ -176,7 +217,6 @@ def set_vocab(gguf_writer: gguf.GGUFWriter, dir_model: Path, hparams: dict): if added_tokens_decoder[i].special or _does_token_look_special(token): toktypes.append(gguf.TokenType.CONTROL) else: - # Pre-normalize user-defined spaces (for Gemma-style tokenizers) token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ") toktypes.append(gguf.TokenType.USER_DEFINED) @@ -191,14 +231,105 @@ def set_vocab(gguf_writer: gguf.GGUFWriter, dir_model: Path, hparams: dict): gguf_writer.add_token_types(toktypes) special_vocab = gguf.SpecialVocab(dir_model, load_merges=True) - # Override EOS token: PyTorch tokenizer appends <|endoftext|> (151643) as the - # sentence-end marker, not <|im_end|> (151645). For last-token pooling to work - # correctly, llama.cpp must append the same token. - special_vocab.special_token_ids["eos"] = 151643 + + if arch == "qwen3": + # Override EOS token: PyTorch tokenizer appends <|endoftext|> (151643) as the + # sentence-end marker, not <|im_end|> (151645). For last-token pooling to work + # correctly, llama.cpp must append the same token. + special_vocab.special_token_ids["eos"] = 151643 + special_vocab.add_to_gguf(gguf_writer) - # Embedding models need EOS token appended for last-token pooling - gguf_writer.add_add_eos_token(True) + if arch == "qwen3": + # Embedding models need EOS token appended for last-token pooling + gguf_writer.add_add_eos_token(True) + + +def _set_vocab_gemma3(gguf_writer: gguf.GGUFWriter, dir_model: Path, + tokenizer, vocab_size: int): + """Set SPM-compatible vocab for Gemma3. + + Gemma's tokenizer is SentencePiece-based (BPE variant with ▁ space prefix + and byte fallback). We read tokenizer.json to extract vocab and compute + BPE merge scores, then write as tokenizer.ggml.model = "llama" so llama.cpp + uses the SPM code path (correct pre-tokenizer behavior for CJK etc.). + + Score assignment: + - BPE merge results get scores derived from merge rank (lower rank = higher score) + - Single-char / byte tokens get score 0 + - Special / added tokens get score -1000 + """ + tokenizer_json_file = dir_model / "tokenizer.json" + if not tokenizer_json_file.exists(): + raise FileNotFoundError(f"tokenizer.json not found in {dir_model}") + + with open(tokenizer_json_file, encoding="utf-8") as f: + tokenizer_json = json.load(f) + + bpe_vocab = tokenizer_json["model"]["vocab"] # token_str -> token_id + bpe_merges = tokenizer_json["model"].get("merges", []) + + # Build merge result -> rank mapping for score computation + # merge_scores[result_token] = -rank (lower rank = earlier merge = higher priority) + merge_scores: dict[str, float] = {} + for rank, merge in enumerate(bpe_merges): + if isinstance(merge, list): + result = "".join(merge) + else: + parts = merge.split(" ", 1) + result = "".join(parts) + if result not in merge_scores: + merge_scores[result] = -float(rank) + + # Build token arrays + reverse_vocab = {v: k for k, v in bpe_vocab.items()} + added_tokens_decoder = tokenizer.added_tokens_decoder + + tokens: list[bytes] = [] + scores: list[float] = [] + toktypes: list[int] = [] + + for i in range(vocab_size): + if i not in reverse_vocab: + tokens.append(f"[PAD{i}]".encode("utf-8")) + scores.append(-10000.0) + toktypes.append(gguf.TokenType.UNUSED) + continue + + token_str = reverse_vocab[i] + token_bytes = token_str.encode("utf-8") + + # Determine token type + if i in added_tokens_decoder: + tok_data = added_tokens_decoder[i] + if tok_data.special or _does_token_look_special(token_str): + toktypes.append(gguf.TokenType.CONTROL) + else: + toktypes.append(gguf.TokenType.USER_DEFINED) + scores.append(-1000.0) + elif token_str.startswith("<0x") and token_str.endswith(">") and len(token_str) == 6: + # Byte token: <0xHH> + toktypes.append(gguf.TokenType.BYTE) + scores.append(0.0) + elif token_str == "": + toktypes.append(gguf.TokenType.UNKNOWN) + scores.append(0.0) + else: + toktypes.append(gguf.TokenType.NORMAL) + # Score from merge rank, or 0 for single-char tokens + scores.append(merge_scores.get(token_str, 0.0)) + + tokens.append(token_bytes) + + gguf_writer.add_tokenizer_model("llama") + gguf_writer.add_tokenizer_pre("default") + gguf_writer.add_token_list(tokens) + gguf_writer.add_token_scores(scores) + gguf_writer.add_token_types(toktypes) + gguf_writer.add_add_space_prefix(False) + + special_vocab = gguf.SpecialVocab(dir_model, load_merges=False) + special_vocab.add_to_gguf(gguf_writer) # --------------------------------------------------------------------------- @@ -260,7 +391,7 @@ def set_gguf_parameters(gguf_writer: gguf.GGUFWriter, hparams: dict, dir_model: pooling_type = gguf.PoolingType.MEAN gguf_writer.add_pooling_type(pooling_type) - logger.info(f" n_layers={n_layers}, n_embd={n_embd}, n_head={n_head}, n_head_kv={n_head_kv}, n_ff={n_ff}") + logger.info(f" n_layers={n_layers}, n_embd={n_embd}, n_head={n_head}, n_head_kv={n_head_kv}, n_ff={n_ff}, head_dim={head_dim}") # --------------------------------------------------------------------------- @@ -366,7 +497,7 @@ def quantize_to_i2_s(w: np.ndarray) -> np.ndarray: # --------------------------------------------------------------------------- def main(): - parser = argparse.ArgumentParser(description="Convert bitnet-embeddings to GGUF") + parser = argparse.ArgumentParser(description="Convert bitnet-embeddings (Qwen3/Gemma3) to GGUF") parser.add_argument("model", type=Path, help="Model directory") parser.add_argument("--outfile", type=Path, default=None, help="Output GGUF file") parser.add_argument("--outtype", choices=["f32", "f16", "i2_s"], default="f16", @@ -390,9 +521,12 @@ def main(): with open(dir_model / "config.json") as f: hparams = json.load(f) - arch = hparams.get("model_type", "qwen3") - assert arch == "qwen3", f"Expected qwen3 architecture, got {arch}" + arch = hparams.get("model_type", "") + if arch not in SUPPORTED_ARCHS: + logger.error(f"Unsupported model_type '{arch}'. Supported: {list(SUPPORTED_ARCHS.keys())}") + sys.exit(1) + gguf_arch = SUPPORTED_ARCHS[arch] n_layers = hparams["num_hidden_layers"] # Determine ftype @@ -403,20 +537,20 @@ def main(): else: # i2_s ftype = 40 # LLAMA_FTYPE_MOSTLY_I2_S - logger.info(f"Converting {dir_model.name} to GGUF ({args.outtype})") + logger.info(f"Converting {dir_model.name} (arch={arch}) to GGUF ({args.outtype})") # Create GGUF writer - gguf_writer = gguf.GGUFWriter(str(args.outfile), "qwen3") + gguf_writer = gguf.GGUFWriter(str(args.outfile), gguf_arch) # Set parameters set_gguf_parameters(gguf_writer, hparams, dir_model, ftype) # Set vocab logger.info("Setting tokenizer/vocab...") - set_vocab(gguf_writer, dir_model, hparams) + set_vocab(gguf_writer, dir_model, hparams, arch) # Build tensor name map - tensor_map = build_tensor_name_map(n_layers) + tensor_map = build_tensor_name_map(n_layers, arch) # Process tensors logger.info("Processing tensors...") @@ -473,6 +607,16 @@ def main(): else: # norms, 1D tensors + # Gemma3 RMSNorm uses (1+w)*x instead of w*x; preprocess w -> w+1 + # so llama.cpp's standard RMSNorm produces correct results. + # NOTE: *_norm_in weights are BitLinear standard RMSNorm (initialized ~1.0), + # NOT Gemma3RMSNorm (initialized ~0.0), so they must NOT get +1. + is_gemma3_native_norm = (arch == "gemma3_text" and is_norm + and not gguf_name.endswith("_norm_in.weight")) + if is_gemma3_native_norm: + data = data.astype(np.float32) + 1.0 + logger.info(f" [Gemma3 norm offset] {gguf_name}: applied w = w + 1") + if args.outtype in ("f16", "i2_s"): data = data.astype(np.float16) logger.info(f" {gguf_name}: {list(data_torch.shape)} {old_dtype} -> float16")