Add files using upload-large-folder tool

README.md

@@ -84,12 +84,14 @@ it invents from the prompt alone. Drama 3 FP8, experts in RAM, temperature 1.0,
 
 ## Files
 
-**Start with FP8**: it holds Fish Audio's own expert weights, byte for byte, next to the checkpoint's bf16 layers. Take
-a smaller file when your system RAM is short of the FP8 file's ~48 GB. Every model file needs the codec.
+**Start with FP8**: it holds Fish Audio's own expert weights, byte for byte, next to the checkpoint's bf16 layers. The
+int8 file is the same size and measured as close to the original. Take a smaller file when your system RAM is short
+of the FP8 file's ~48 GB. Every model file needs the codec.
 
 | File | Size | Experts | System RAM | Notes |
 |---|---|---|---|---|
 | [`drama3_fp8experts_bf16.safetensors`](drama3_fp8experts_bf16.safetensors) | 38.6 GB | FP8, as Fish Audio ships them | ~48 GB | the original weights |
+| [`drama3_int8experts_bf16.safetensors`](drama3_int8experts_bf16.safetensors) | 38.7 GB | int8 ConvRot | ~48 GB | no silent take in 12 cloned test takes |
 | [`drama3_w6a8experts_bf16.safetensors`](drama3_w6a8experts_bf16.safetensors) | 32.7 GB | W6A8 | ~40 GB | no silent take in 16 cloned test takes |
 | [`drama3_w4a8experts_bf16.safetensors`](drama3_w4a8experts_bf16.safetensors) | 24.6 GB | W4A8 | ~32 GB | falls silent more often with real voice clips |
 | [`drama3_codec.safetensors`](drama3_codec.safetensors) | 1.6 GB | | | the original codec, fp32 |
@@ -125,6 +127,7 @@ voice, ~9 s of speech, prompt pass included, measured on 2026-10-14:
 | GPU | File | Expert cache | ms per frame |
 |---|---|---|---|
 | RTX 4090 (PCIe 4.0 ×8) | W4A8 | 16 GB | 53 |
+| RTX 4090 (PCIe 4.0 ×8) | int8 | 16 GB | 72 |
 | RTX 3090 (PCIe 4.0 ×8) | FP8 | 16 GB | 83 |
 | RTX 3090 held to 12 GB | FP8 | automatic | 137 |
 | RTX 3090 held to 8 GB | FP8 | automatic | 214 |
@@ -134,14 +137,15 @@ voice, ~9 s of speech, prompt pass included, measured on 2026-10-14:
 From [fishaudio/drama-3](https://huggingface.co/fishaudio/drama-3) with the pack's `tools/convert.py`:
 
 - **FP8**: the routed experts' FP8 bytes and 128×128 block scales copied unchanged, gate and up projections fused.
-- **W6A8 / W4A8**: the experts dequantized from FP8 in fp32 and requantized (6-bit or 4-bit codebook weights with
-  int8 activations, ComfyUI's comfy_kitchen formats); relative weight error 2.1% and 7.3%.
+- **int8 / W6A8 / W4A8**: the experts dequantized from FP8 in fp32 and requantized, in ComfyUI's comfy_kitchen
+  formats: int8 ConvRot (per-channel int8 after a rotation in groups of 256), or 6-bit or 4-bit codebook weights with
+  int8 activations; relative weight error 0.83%, 2.1% and 7.3%.
 - **Every file**: the attention, DeltaNet, shared-expert and codebook-decoder layers are the checkpoint's own bf16;
   the LM head keeps only the rows generation samples from (end of speech and the 4096 semantic codes).
 - **Codec**: Fish Audio's S2-Pro codec as shipped with Drama 3, weight norm folded in, fp32.
 
 Against Fish Audio's own PyTorch code, with the same frames forced, the main token's distribution differs by a KL of
-0.004–0.007 nats with the FP8 file, the noise of bf16 arithmetic.
+0.004–0.008 nats with the FP8 and int8 files, the noise of bf16 arithmetic.
 
 ## License
 

drama3_int8experts_bf16.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f407996b17583ba857ace14e0330a13788cf89f4d47cfdcd4766ad5106576899
+size 38696401232