Add Llama 3.2 3B affine-Q4 cellm model

Adds the verified 2.37 GiB affine-Q4 v1 artifact, tokenizer files, model card, and conversion manifest.

.gitattributes

@@ -70,3 +70,4 @@ nanowhale-100m-v1/nanowhale-100m-v1.cellm filter=lfs diff=lfs merge=lfs -text
 LFM2.5-230M/LFM2.5-230M-int4.cellm filter=lfs diff=lfs merge=lfs -text
 LFM2.5-230M/LFM2.5-230M.cellm filter=lfs diff=lfs merge=lfs -text
 LFM2.5-230M/LFM2.5-230M-int4-v2.cellm filter=lfs diff=lfs merge=lfs -text
+Llama-3.2-3B-Instruct-Q4-affine-v1/Llama-3.2-3B-Instruct-Q4-affine-v1.cellm filter=lfs diff=lfs merge=lfs -text

Llama-3.2-3B-Instruct-Q4-affine-v1/Llama-3.2-3B-Instruct-Q4-affine-v1.cellm

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f408ba787161090b5a9e0aa02248655fd81d8f256f8198cc767add9f3a8ffcad
+size 2550044032

Llama-3.2-3B-Instruct-Q4-affine-v1/README.md

@@ -0,0 +1,59 @@
+---
+library_name: cellm
+base_model: meta-llama/Llama-3.2-3B-Instruct
+license: llama3.2
+tags:
+  - cellm
+  - llama
+  - quantized
+  - affine-q4
+  - cpu
+---
+
+# Llama 3.2 3B Instruct — cellm affine-Q4 v1
+
+A `cellm` conversion of Llama 3.2 3B Instruct for local inference. The model preserves unsigned affine 4-bit weights with per-64-value FP32 scales and biases instead of symmetrically requantizing them.
+
+## Artifact
+
+- File: `Llama-3.2-3B-Instruct-Q4-affine-v1.cellm`
+- Size: `2,550,044,032` bytes (2.37 GiB)
+- SHA-256: `f408ba787161090b5a9e0aa02248655fd81d8f256f8198cc767add9f3a8ffcad`
+- Layers: 28
+- Quantization: unsigned affine Q4, group size 64
+- RoPE: Llama 3 scaled interleaved RoPE
+- KV cache used for validation: F16
+
+See `manifest.json` for conversion provenance, source hashes, tensor counts, model configuration, and runtime baseline metadata.
+
+## CPU inference
+
+```sh
+./target/release/infer \
+  --model Llama-3.2-3B-Instruct-Q4-affine-v1.cellm \
+  --tokenizer tokenizer.json \
+  --prompt "What is the capital of France?" \
+  --chat --gen 32 --temperature 0 --top-k 1 \
+  --repeat-penalty 1.0 --backend cpu --kv-encoding f16
+```
+
+Expected deterministic response:
+
+```text
+The capital of France is Paris.
+```
+
+## Requirements
+
+Use a `cellm` runtime that supports:
+
+- packed unsigned affine-Q4 weights (`u32` tensor convention)
+- per-group FP32 `.scales` and `.biases`
+- Llama 3 interleaved scaled RoPE
+- Llama 3 chat formatting and `<|eot_id|>` stopping
+
+Repository: https://github.com/jeffasante/cellm
+
+## License
+
+The model is derived from `meta-llama/Llama-3.2-3B-Instruct` and remains subject to the Llama 3.2 Community License and Acceptable Use Policy. Users are responsible for complying with the upstream terms.

Llama-3.2-3B-Instruct-Q4-affine-v1/manifest.json

@@ -0,0 +1,84 @@
+{
+  "schema_version": 1,
+  "baseline_id": "llama-3.2-3b-instruct-q4-affine-v1",
+  "artifact": {
+    "path": "models/Llama-3.2-3B-Instruct-Q4-affine-v1.cellm",
+    "sha256": "f408ba787161090b5a9e0aa02248655fd81d8f256f8198cc767add9f3a8ffcad",
+    "expected_file_size": 2550044032,
+    "format": "cellm",
+    "format_version": 1,
+    "tensor_count": 646,
+    "tensor_dtype_counts": {
+      "f16": 58,
+      "f32": 392,
+      "u32": 196
+    }
+  },
+  "source_model": {
+    "path": "models/gguf/Llama-3.2-3B-Instruct-Q4.base",
+    "sha256": "20c803b2eeee70a7dada67ca0ee2d75081d95d0e35bac27d1857f9e3fdd84615",
+    "expected_file_size": 2380601344,
+    "format": "base",
+    "format_version": 1,
+    "source_checkpoint": "src-Llama-3.2-3B-Instruct"
+  },
+  "tokenizer": {
+    "path": "models/Llama-3.2-3B-Instruct/tokenizer.json",
+    "sha256": "6b918bc5f3f1c33371d5ac45a53d565a737a7ff1bdf3a8b8055e09eec6a25ff4",
+    "config_path": "models/Llama-3.2-3B-Instruct/tokenizer_config.json",
+    "config_sha256": "00e9df46f6d2f18d1e44fa2741634f04603ec16a64b5062f7505ebf1700880ff",
+    "chat_template": "llama3",
+    "chat_template_markers": [
+      "<|start_header_id|>",
+      "<|end_header_id|>",
+      "<|eot_id|>"
+    ]
+  },
+  "conversion": {
+    "timestamp_utc": "2026-07-22T23:43:30Z",
+    "command": "python3 tools/convert_base_to_cellm.py --input models/gguf/Llama-3.2-3B-Instruct-Q4.base --output models/Llama-3.2-3B-Instruct-Q4-affine-v1.cellm --tokenizer models/Llama-3.2-3B-Instruct --quantize affine-i4",
+    "quantization_scheme": "unsigned-affine-i4",
+    "group_size": 64,
+    "dequantization_formula": "value = q_unsigned * scale + bias",
+    "converter_path": "tools/convert_base_to_cellm.py",
+    "converter_sha256": "0f5395839bc9f8136de6913f99cf09059cb15e48b840ed52490ac6813b521d79",
+        "converter_commit": "fd47a71b7cafd102a0fc31003a609225a378cb5a",
+    "working_tree_dirty": true
+  },
+  "runtime": {
+    "commit": "fd47a71b7cafd102a0fc31003a609225a378cb5a",
+    "working_tree_dirty": true,
+    "source_sha256": {
+      "crates/cellm-model/src/llama.rs": "4a57ac121435f6298e32a9363e569029d6d917ec9a2d48be8715f7175a86132d",
+      "crates/cellm-model/src/cellm_file.rs": "4e5697a2c6f77cf867ee4200da2eea040e91c38b9cc0a47140e45aba9757e50d",
+      "crates/cellm-kernels/src/cpu_kernels.rs": "ac8e530e9b46ed398edea0090d46cfccf45f3f9cc0e0ddaffd9dc4f83bdfef08",
+      "tools/infer/src/main.rs": "bbe18febed0da38bc82f786ca7b91409cccf8cb66f34e07c1c534e24606c108c"
+    }
+  },
+  "model_config": {
+    "architecture": "llama",
+    "layers": 28,
+    "hidden_size": 3072,
+    "intermediate_size": 8192,
+    "attention_heads": 24,
+    "kv_heads": 8,
+    "head_dim": 128,
+    "vocab_size": 128256,
+    "rope_theta": 500000.0,
+    "rope_mode": "interleaved",
+    "rope_scaling": {
+      "type": "llama3",
+      "factor": 32.0,
+      "original_max_position_embeddings": 8192,
+      "low_freq_factor": 1.0,
+      "high_freq_factor": 4.0
+    },
+    "bos_token_id": 128000,
+    "eos_token_id": 128001,
+    "stop_token_ids": {
+      "eom_id": 128008,
+      "eot_id": 128009
+    }
+  },
+  "oracle_fixture": "tests/baselines/llama_3_2_3b_affine_v1_oracle.json"
+}

Llama-3.2-3B-Instruct-Q4-affine-v1/tokenizer_config.json

@@ -0,0 +1,3 @@
+{
+  "chat_template": "{{- bos_token }}\n{%- if custom_tools is defined %}\n    {%- set tools = custom_tools %}\n{%- endif %}\n{%- if not tools_in_user_message is defined %}\n    {%- set tools_in_user_message = true %}\n{%- endif %}\n{%- if not date_string is defined %}\n    {%- if strftime_now is defined %}\n        {%- set date_string = strftime_now(\"%d %b %Y\") %}\n    {%- else %}\n        {%- set date_string = \"26 Jul 2024\" %}\n    {%- endif %}\n{%- endif %}\n{%- if not tools is defined %}\n    {%- set tools = none %}\n{%- endif %}\n\n{#- This block extracts the system message, so we can slot it into the right place. #}\n{%- if messages[0]['role'] == 'system' %}\n    {%- set system_message = messages[0]['content']|trim %}\n    {%- set messages = messages[1:] %}\n{%- else %}\n    {%- set system_message = \"\" %}\n{%- endif %}\n\n{#- System message #}\n{{- \"<|start_header_id|>system<|end_header_id|>\\n\\n\" }}\n{%- if tools is not none %}\n    {{- \"Environment: ipython\\n\" }}\n{%- endif %}\n{{- \"Cutting Knowledge Date: December 2023\\n\" }}\n{{- \"Today Date: \" + date_string + \"\\n\\n\" }}\n{%- if tools is not none and not tools_in_user_message %}\n    {{- \"You have access to the following functions. To call a function, please respond with JSON for a function call.\" }}\n    {{- 'Respond in the format {\"name\": function name, \"parameters\": dictionary of argument name and its value}.' }}\n    {{- \"Do not use variables.\\n\\n\" }}\n    {%- for t in tools %}\n        {{- t | tojson(indent=4) }}\n        {{- \"\\n\\n\" }}\n    {%- endfor %}\n{%- endif %}\n{{- system_message }}\n{{- \"<|eot_id|>\" }}\n\n{#- Custom tools are passed in a user message with some extra guidance #}\n{%- if tools_in_user_message and not tools is none %}\n    {#- Extract the first user message so we can plug it in here #}\n    {%- if messages | length != 0 %}\n        {%- set first_user_message = messages[0]['content']|trim %}\n        {%- set messages = messages[1:] %}\n    {%- else %}\n        {{- raise_exception(\"Cannot put tools in the first user message when there's no first user message!\") }}\n{%- endif %}\n    {{- '<|start_header_id|>user<|end_header_id|>\\n\\n' -}}\n    {{- \"Given the following functions, please respond with a JSON for a function call \" }}\n    {{- \"with its proper arguments that best answers the given prompt.\\n\\n\" }}\n    {{- 'Respond in the format {\"name\": function name, \"parameters\": dictionary of argument name and its value}.' }}\n    {{- \"Do not use variables.\\n\\n\" }}\n    {%- for t in tools %}\n        {{- t | tojson(indent=4) }}\n        {{- \"\\n\\n\" }}\n    {%- endfor %}\n    {{- first_user_message + \"<|eot_id|>\"}}\n{%- endif %}\n\n{%- for message in messages %}\n    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}\n        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\\n\\n'+ message['content'] | trim + '<|eot_id|>' }}\n    {%- elif 'tool_calls' in message %}\n        {%- if not message.tool_calls|length == 1 %}\n            {{- raise_exception(\"This model only supports single tool-calls at once!\") }}\n        {%- endif %}\n        {%- set tool_call = message.tool_calls[0].function %}\n        {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' -}}\n        {{- '{\"name\": \"' + tool_call.name + '\", ' }}\n        {{- '\"parameters\": ' }}\n        {{- tool_call.arguments | tojson }}\n        {{- \"}\" }}\n        {{- \"<|eot_id|>\" }}\n    {%- elif message.role == \"tool\" or message.role == \"ipython\" %}\n        {{- \"<|start_header_id|>ipython<|end_header_id|>\\n\\n\" }}\n        {%- if message.content is mapping or message.content is iterable %}\n            {{- message.content | tojson }}\n        {%- else %}\n            {{- message.content }}\n        {%- endif %}\n        {{- \"<|eot_id|>\" }}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' }}\n{%- endif %}\n"
+}