NAMING.md

# Sealed naming standard — coherent AMD family

## Hard constraint from ROCmFPX types

| Wanted name | Exists? | Notes |
|-------------|---------|-------|
| `Q4_0_ROCMFP4_COHERENT` + imatrix | **YES** | Preferred dual-GPU default |
| `Q5_*_ROCMFP4_COHERENT` | **NO** | No ROCmFP4 Q5 type in this fork |
| `Q6_0_ROCMFP4_COHERENT` | **NO** | Q6 is **ROCmFPX**, not ROCmFP4 |
| `Q6_0_ROCMFPX_AGENT` + imatrix | **YES** | Q6 “coherent / agent” profile |

**ROCMFP4_COHERENT is a Q4-only quant type.**  
Higher bit “coherent” uses **ROCMFPX_AGENT**, not the string `ROCMFP4_COHERENT`.

## Shipped set (this dir)

| File | Role |
|------|------|
| `Qwen3.6-35B-A3B-Q4_0_ROCMFP4_COHERENT-imatrix.gguf` | Default — format ROCmFP4 + profile COHERENT + imatrix (~19G) |
| `Qwen3.6-35B-A3B-Q6_0_ROCMFPX_AGENT-imatrix.gguf` | High quality AMD coherent Q6 (~31G, tight on 32GB) |
| `Qwen3.6-35B-A3B-imatrix.gguf` | Shared imatrix |

**Removed (not coherent AMD):**
- `Q5_K_M-imatrix` (stock Q5, not ROCmFP4/ROCmFPX coherent)
- `Q6_0_ROCMFPX-imatrix` (base Q6, not AGENT)

## Filename pattern (locked)

```text
Qwen3.6-35B-A3B-{TYPE}-imatrix.gguf
```

Where `{TYPE}` is the **exact llama-quantize ftype string**:
- `Q4_0_ROCMFP4_COHERENT`
- `Q6_0_ROCMFPX_AGENT`
- (future) `Q8_0_ROCMFPX_AGENT`, `Q3_0_ROCMFPX_AGENT`, etc.

Do **not** invent `Q5_0_ROCMFP4_COHERENT` or rename stock quants to look like ROCmFP4.

## Optional future

| Type | Size (est.) | When |
|------|-------------|------|
| `Q6_0_ROCMFPX_AGENT_LEAN` | ~27G | If full AGENT is too tight on 32GB |
| `Q8_0_ROCMFPX_AGENT` | ~34G | Dual-GPU / short ctx only |