{ "schema": "qwen38-dflash2-rtx5090-qualification/v1", "observed_at": "2026-08-21", "completion_state": "complete_with_rejection_and_restoration", "evidence_labels": [ "functional", "compatibility-only", "capacity" ], "decision_labels": [ "rejected", "no-promotion" ], "topology": { "measured_host": "Fakoli Mid Mod qualification lane", "measured_gpu": "NVIDIA GeForce RTX 5090, 32,607 MiB, sm_120", "runtime_environment": "Windows 11, Docker Desktop, WSL2", "protected_or_co_resident_gpu": "none", "external_router_policy_changed": false }, "identity": { "target_repository": "RadixArk/Qwen3.8-27B-NVFP4", "target_revision": "554ebba9b5f1b79dc11246341960360e6ef05ef4", "draft_repository": "incoai/Qwen3.8-27B-DFlash2", "draft_revision": "dedf8df68adfb1afeaf7b7480c0a0243108177b4", "runtime_image": "lmsysorg/sglang:dev@sha256:8acc563e39f4e79118cc3c11cb5a8893ca8da140b2280cdd24a9f3bfe38835a0", "runtime_source_revision": "f825d729363136a2d4a4b330fa694d0b37a878fa", "runtime_image_version": "lmsysorg/sglang:nightly-dev-20260821-f825d729", "served_model_name": "qwen38-27b-radixark-nvfp4-sglang-rtx5090-128k-mm" }, "recipe": { "selector": "RTX 5090 / Default / NVFP4 / Single Node / DFLASH2 / High-Throughput / float32", "configured_context_tokens": 262144, "max_running_requests": 1, "cuda_graph_max_batch_size": 1, "kv_cache_dtype": "fp8_e4m3", "attention_backend": "flashinfer", "speculative_algorithm": "DFLASH", "speculative_num_draft_tokens": 8, "mamba_full_memory_ratio": 10, "mamba_radix_cache_strategy": "extra_buffer_lazy", "mamba_ssm_dtype": "float32", "thinking_default": "disabled", "local_contract_additions": [ "immutable target and draft snapshot paths", "stable upstream served-model name", "explicit 262,144 configured context", "CPU multimodal feature transport", "thinking-disabled server default" ] }, "attempts": [ { "attempt": 1, "mem_fraction_static": 0.895, "classification": "local transcription error", "outcome": "startup_failed", "failure": "Hybrid state cache was too small to serve any requests: max_mamba_cache_size=2, four state slots per request, max_num_reqs=0.", "restoration": "Exact 128K baseline restored healthy and passed the complete preflight." }, { "attempt": 2, "mem_fraction_static": 0.945, "classification": "official selector with local contract additions", "outcome": "started_then_rejected_by_capacity_gate", "startup": { "target_weight_memory_gb": 20.14, "draft_weight_memory_gb": 3.92, "max_mamba_cache_size": 5, "target_kv_pool_tokens": 24347, "effective_max_input_tokens": 24341, "draft_cuda_graph": "disabled because zero free GPU memory remained after target initialization" }, "preflight": { "smoke_http_status": 200, "structured_json_http_status": 200, "tool_batch": "20/20 HTTP 200 responses", "long_context": "failed", "long_context_prompt_tokens": 105649, "long_context_error": "Input length exceeds the maximum allowed length of 24,341 tokens." }, "dflash_log_samples": [ { "accept_length": 5.05, "accept_rate": 0.58 }, { "accept_length": 5.60, "accept_rate": 0.66 } ] }, { "attempt": 3, "classification": "BF16 state-width diagnostic at the official BF16 memory fraction", "changes_from_prior": [ "mamba_ssm_dtype=bfloat16", "mem_fraction_static=0.90", "max_mamba_cache_size=4" ], "outcome": "started", "target_kv_pool_tokens": 30984, "available_gpu_memory_after_graphs_gb": 0.91 }, { "attempt": 4, "classification": "one-variable memory-fraction diagnostic", "changes_from_prior": [ "mem_fraction_static=0.945" ], "outcome": "started", "target_kv_pool_tokens": 64790, "target_prefill_cuda_graph_memory_gb": 1.13, "available_gpu_memory_after_target_prefill_graph_gb": 0.0 }, { "attempt": 5, "classification": "one-variable prefill-graph diagnostic", "changes_from_prior": [ "disable_prefill_cuda_graph=true" ], "outcome": "started", "target_kv_pool_tokens": 64790, "available_gpu_memory_after_verify_graph_gb": 1.0, "finding": "Disabling prefill CUDA graphs restored runtime headroom but did not resize the KV pool because KV allocation precedes graph capture." }, { "attempt": 6, "classification": "best safe single-request DFlash2 capacity arm", "changes_from_prior": [ "disable_radix_cache=true", "max_mamba_cache_size=1" ], "outcome": "functional_pass_below_route_capacity", "target_kv_pool_tokens": 70262, "mamba_state": { "persistent_ssm_state_gb": 0.14, "intermediate_ssm_state_gb": 1.12, "intermediate_conv_window_gb": 0.02 }, "preflight": { "short_coding": "pass", "structured_json": "pass", "needle_generated_context": 60000, "needle_actual_prompt_tokens": 49549, "needle": "pass", "tool_batch": "20/20 clean" }, "dflash_log_sample": { "accept_length": 5.6, "accept_rate": 0.66, "generation_tokens_per_second": 45.63 } } ], "diagnosis": { "root_cause": "The checkpoint and DFlash implementation are compatible; the failure is resident-memory capacity on a 32 GB GPU.", "fixed_weight_memory_gb": 24.06, "configured_context_is_not_resident_capacity": true, "best_safe_dflash_kv_pool_tokens": 70262, "retained_route_prompt_gate_tokens": 105649, "primary_constraint": "The eight-token DFlash verification path retains a 1.12 GB intermediate BF16 SSM state cache in addition to target and draft weights, persistent state, KV pools, and runtime graphs." }, "recommended_adjustments": { "distinct_short_context_candidate": [ "mamba_ssm_dtype=bfloat16", "mem_fraction_static=0.945", "max_mamba_cache_size=1", "disable_radix_cache=true", "disable_prefill_cuda_graph=true", "use a distinct truthful served name and admission limit" ], "existing_128k_route": "Retain the qualified no-spec recipe.", "future_speculative_128k_research": "Evaluate an EAGLE/MTP replay-SSM path documented to avoid resident speculative verification states, or use more aggregate VRAM; do not extrapolate that option to DFlash without qualification." }, "restoration": { "recipe": "RadixArk/Qwen3.8-27B-NVFP4-RTX5090-128K", "runtime_image": "lmsysorg/sglang:qwen38-27b@sha256:506525a5907ea22c9d445afb7c03603959b912de034d86915cf17da814f1a124", "status": "healthy", "preflight": { "short_coding": "pass", "structured_json": "pass", "needle_context": "pass at approximately 128,000 generated context units and 105,649 actual prompt tokens", "needle_marker": "ZEBRA-42917-QUARTZ", "tool_batch": "20/20 clean" } }, "limitations": [ "This was not an eval benchmark artifact and does not establish stable throughput or quality.", "DFlash acceptance figures are scheduler-log samples, not a controlled performance comparison.", "No routed request was used as public evidence; the external router policy was left unchanged.", "The configured 262,144-token context is not the measured request capacity of this memory layout.", "The 70,262-token tuned arm is a short-context lead, not a promoted or route-qualified service." ], "sources": [ "https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8-27B", "https://github.com/sgl-project/sglang/blob/f825d729363136a2d4a4b330fa694d0b37a878fa/docs/src/snippets/configs/Qwen/qwen3.8-27b.jsx", "https://github.com/sgl-project/sglang/commit/f825d729363136a2d4a4b330fa694d0b37a878fa", "https://huggingface.co/incoai/Qwen3.8-27B-DFlash2/tree/dedf8df68adfb1afeaf7b7480c0a0243108177b4" ] }