dflash A/B: attention_backend FLASHINFER→TRITON_ATTN — FLASHINFER rejects fp8 kv-cache-dtype on this build; TRITON_ATTN is what the kindling lane uses for dflash

recipes/glm53-flash-fp4local-2x-dflash/recipe.toml

@@ -63,7 +63,7 @@ args = [
     "--enable-auto-tool-choice",
     "--gpu-memory-utilization", "0.87",
     "--speculative-config", """
-'{"method":"dflash","model":"{{dflash2}}","num_speculative_tokens":7,"rejection_sample_method":"standard","attention_backend":"FLASHINFER","draft_load_config":{"load_format":"auto"}}'""",
+'{"method":"dflash","model":"{{dflash2}}","num_speculative_tokens":7,"rejection_sample_method":"standard","attention_backend":"TRITON_ATTN","draft_load_config":{"load_format":"auto"}}'""",
 ]
 
 # DFlash2 drafter (A/B vs the card's embedded MTP head): distilled against the