Update README for corrected quantization (527 layers, embeddings quantized)

README.md

@@ -17,9 +17,9 @@ tags:
 # Gemma 4 E2B SnowFox MLX 6-bit (affine, group 64)
 
 Standard MLX-VLM 6-bit affine weight quantization of the SnowFox model —
-the MLX equivalent of GGUF `Q6_K`. A genuine MLX-VLM package (quantized
-`safetensors` + `config.json` carrying a `quantization` field), not a GGUF
-file or a renamed HF checkpoint.
+the MLX equivalent of GGUF `Q6_K`. This is a genuine MLX-VLM package
+(quantized `safetensors` + `config.json` carrying a `quantization` field),
+not a GGUF file or a renamed HF checkpoint.
 
 SnowFox is a language-only LoRA merge based on Google's Gemma 4 E2B
 instruction QAT-derived checkpoint. The image and audio towers were frozen
@@ -35,18 +35,20 @@ tokenizer needed by MLX-VLM.
 
 ## What is quantized
 
-- **525 linear layers** (`q/k/v/o` projections, MLP gate/up/down, multimodal
-  projectors) are 6-bit affine quantized: packed `uint8` `weight` (4 values per
-  3 bytes) + float16 `scales`/`biases`.
-- **Embeddings, layer norms, convolutions, and biases stay float16** — matching
-  MLX-VLM's standard Linear-only affine quantization. The large per-layer input
-  embedding (`embed_tokens_per_layer`) is retained at full precision, so the
-  model needs ~7 GB unified memory on Apple Silicon.
+- **527 linear and embedding layers** (`q/k/v/o` projections, MLP gate/up/down,
+  multimodal projectors, and the large embeddings) are 6-bit affine quantized:
+  packed `uint32` `weight` (4 values per 3 bytes, low bits first) + float16
+  `scales`/`biases`.
+- **Layer norms, convolutions, and biases stay float16** — matching MLX-VLM's
+  standard affine quantization. The dense per-layer input embedding
+  (`embed_tokens_per_layer`) **is** quantized here, so the package stays under
+  ~4.2 GB rather than the ~7 GB a dense embedding would force.
 
 ## Package contents
 
-- `model-00001-of-00003.safetensors` … `model-00003-of-00003.safetensors`: the
-  one 6-bit MLX model (~7.4 GB total).
+- `model-00001-of-00002.safetensors` (4,000,099,030 bytes) and
+  `model-00002-of-00002.safetensors` (166,857,424 bytes): the 6-bit MLX model
+  (~4.2 GB total).
 - `model.safetensors.index.json`: complete shard map.
 - `config.json` (with `quantization` + `quantization_config`), `generation_config.json`,
   `processor_config.json`, tokenizer files, and `chat_template.jinja`.
@@ -56,14 +58,14 @@ tokenizer needed by MLX-VLM.
 The conversion host has no Apple-Silicon MLX runtime, so the quantized package
 was structurally validated before upload:
 
-- 1,951 source tensors mapped with no missing or extra keys; 525 linear layers
-  quantized; embeddings left dense.
-- Quantized weight format matches the MLX 6-bit affine contract: 4 values
-  packed per 3 `uint8` bytes (value *k* at bits `[6k, 6k+6)`), dequantization
-  `scale * q + bias`, group 64.
-- Round-trip dequantization of sampled layers (language + audio towers) reproduces
-  the source weights to within 6-bit precision (max absolute error ≤ ~0.004).
-- Config matches MLX-VLM `convert --quantize --q-bits 6` output.
+- 1,951 source tensors mapped with no missing or extra keys; 527 linear +
+  embedding layers quantized.
+- Quantized weight format matches the MLX affine contract: 6-bit values packed
+  4-per-3-bytes (24-bit little-endian word), dequantization `scale * q + bias`,
+  group 64.
+- Round-trip dequantization of sampled layers (attention projections + the
+  2.35B-param `embed_tokens_per_layer`) reproduces the source weights to within
+  6-bit precision (max relative error ≈ 1.4%).
 
 **Apple-Silicon MLX-VLM inference has not been run.** Treat this as a
 structurally validated quantization pending a real Apple-Silicon text / image /