[Bug] Workload blob path format inconsistent

[wenyi-li]

Summary

The workload JSONL files reference blob safetensors using a flat path format, while add_workload_blob_tensor in flashinfer-bench writes blobs using a subdirectory format. This mismatch causes FileNotFoundError when running benchmarks.

Path Format Comparison

1. Workload JSONL (current format - flat):

./blob/workloads/{op_type}/{def_name}_{workload_uuid}.safetensors

Example: ./blob/workloads/gdn/gdn_prefill_qk4_v8_d128_k_last_de7af853-9c30-4e68-a0b7-6c6cd001b0ca.safetensors

2. add_workload_blob_tensor (subdirectory format):

# flashinfer_bench/data/trace_set.py:424-430
file_path = (
    self.blob_workloads_path
    / op_type
    / def_name
    / f"{def_name}_{workload_uuid}.safetensors"
)

This produces:

blob/workloads/{op_type}/{def_name}/{def_name}_{workload_uuid}.safetensors

Example: blob/workloads/gdn/gdn_prefill_qk4_v8_d128_k_last/gdn_prefill_qk4_v8_d128_k_last_de7af853....safetensors

Impact

• Benchmarks fail when loading workloads that reference safetensors inputs
• Error: No such file or directory: .../blob/workloads/gdn/gdn_prefill_qk4_v8_d128_k_last_de7af853....safetensors
• Workers crash during baseline (run_ref) → No healthy workers available

Affected Workloads

• GDN workloads (gdn_prefill_qk4_v8_d128_k_last, gdn_decode_*, etc.)