ggml_validate_row_data rejects the ROCmFPx types it just produced — no FPx recipe can be quantized

[walcz-de]

Summary

llama-quantize offers the ROCmFPx recipes but cannot produce them: the row-data validation rejects the very types it just wrote, so every FPx quantization aborts.

Reproduction

llama-quantize model-f16.gguf out.gguf Q6_0_ROCMFPX_AGENT

ggml_validate_row_data: invalid type 103
... (repeated per tensor)
llama_model_quantize: failed to quantize: quantized data validation failed
llama_quantize: failed to quantize model from '.../model-f16.gguf'

Exit code 1, output file truncated at ~11 MB.

Cause

ggml_validate_row_data() in ggml/src/ggml-quants.c only handles the two FP4 types:

case GGML_TYPE_Q4_0_ROCMFP4:        // 100
case GGML_TYPE_Q4_0_ROCMFP4_FAST:   // 101

The FPx family is not covered:

Q6_0_ROCMFPX = 102
Q8_0_ROCMFPX = 103
Q3_0_ROCMFPX = 104
Q2_0_ROCMFPX = 107

Commit e5b4bc02 ("rocmfp4: validate the FP4 row data instead of rejecting it as unknown") added the two FP4 cases; the FPx types were not carried along.

Q6_0_ROCMFPX_AGENT mixes Q8_0_ROCMFPX (103) into the sensitive tensors — which is exactly what makes it the agent/tool-call recipe — so it trips the gap immediately.

Scope

All recipes in the FPx family are affected, which includes every *_AGENT variant advertised in the quantize type list:

Q3_0_ROCMFPX        Q3_0_ROCMFPX_AGENT
Q6_0_ROCMFPX        Q6_0_ROCMFPX_AGENT      Q6_0_ROCMFPX_LEAN
Q8_0_ROCMFPX        Q8_0_ROCMFPX_AGENT      Q6_0_ROCMFPX_AGENT_LEAN
Q2_0_ROCMFPX

They are selectable and documented, and none of them can be produced.

Fix

Extend ggml_validate_row_data() with the FPx types, following the shape of the two FP4 cases already there.