narandill/vllm repository overview

=== 1. Environment ===
  INFO : GPU          : Orin (nvgpu)
  INFO : Compute cap  : 8.7
  PASS : Compute capability is 8.7 (Jetson Orin / Ampere)
  INFO : Driver       : 540.4.0
  PASS : Driver 540.4.0 identified (Tegra 540.x series, CUDA 12.6 JIT limit)
  INFO : nvcc         : 12.9
  INFO : Python       : Python 3.12.13
  INFO : vLLM         : 0.21.0
  PASS : vLLM 0.21.0 is installed

=== 2. Flash-attention .so cubin targets ===
  INFO : _vllm_fa2_C cubins     : [sm_87 ]
  INFO : _vllm_fa2_C ptx_ver    : [none]
  PASS : _vllm_fa2_C contains native sm_87 cubin
  PASS : _vllm_fa2_C has no PTX blob (correct — no JIT needed on driver 540.x)
  INFO : _vllm_fa3_C cubins     : [sm_87 ]
  PASS : _vllm_fa3_C built for sm_87 (modern vLLM ≥0.19 — FA3 usable on Jetson)

=== 3. Main extension .so cubin targets ===
  INFO : _C    (main ops) cubins  : [sm_87 ]
  PASS : _C    (main ops) contains native sm_87 cubin
  INFO : _moe_C (MoE ops) cubins  : [sm_87 ]
  PASS : _moe_C (MoE ops) contains native sm_87 cubin

=== 4. vLLM attention backend selection ===
  INFO : compute_capability=8.7  sm=87
backend_selection=FlashAttention (FA2)  [sm_87 is supported, FA3 is supported sm_87]
  PASS : PyTorch confirms compute capability 8.7

=== 5. Quantisation readiness (sm_87 compatible) ===
  INFO : Quant stack      : bn=1|awq=0|gguf=1|ctens=1|methods=empty
  INFO : AWQ partially detected (Transformers-version-dependent)
  PASS : GGUF reader present — community quant packs accessible
  PASS : compressed-tensors installed — generic quant wrappers available
  INFO : FP8 KV-cache     : via_argutils
  PASS : FP8 KV-cache dtype recognised — halves KV VRAM (0.21.0)
  INFO : Launch tip: --kv-cache-dtype fp8  (attention still BF16-computed on Ampere)

=== 6. Speculative-decode readiness ===
  INFO : Spec-Decoders    : draft_model,eagle,gemma4_mtp,medusa,ngram,ngram_gpu (conditional: dflash)
  PASS : Active spec-decoders for sm_87: draft_model,eagle,gemma4_mtp,medusa,ngram,ngram_gpu
  WARN : Conditional spec-decoders (need non-standard attn/backend cfg): dflash
  INFO : Tip: DFlash requires causal=False attention — set --attention-backend triton

=== 7. Runtime / deployment knobs ===
  INFO : Runtime profile  : arg_count=201&engine_ok=True&has_guided_decoding=False&has_opentelemetry=True&has_perf_mode=True&has_vllm__C=True
  PASS : EngineArgs introspected — 201 configurable params surfaced
  PASS : --performance-mode knob available — steer towards interactivity or throughput
  INFO : Guided-decoding arg not surfaced in EngineArgs
  PASS : OpenTelemetry trace collector bundled — observability ON by default
  PASS : vllm._C loads cleanly — JIT monitoring infrastructure likely present

=== Final tally ===

RESULT: PASSED — wheel ready for Jetson Orin SM 8.7