use_fp8_dispatch requires the ElasticBuffer to receive FP8 input.
In production, this is ensured by pre-quantizing via
moe_kernel_quantize_input when is_block_quantized=True.
The test was parametrizing use_fp8_dispatch independently of dtype,
allowing bf16 input with use_fp8_dispatch=True which triggers a
buffer size assertion in DeepEP v2.
Fix:
- Derive use_fp8_dispatch from dtype (True only for FP8 weights)
- Add block_shape=[128, 128] to quant config for FP8 to enable
the block quantization path that pre-quantizes input
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
