Publish promoted generation-3 DFlash2 head

Private promoted checkpoint with paired acceptance and parity metrics; trace data intentionally excluded.

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/dflash2-figure.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,92 @@
+---
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+base_model:
+  - Qwen/Qwen3.8-27B
+inference: false
+tags:
+  - dflash2
+  - speculative-decoding
+  - block-diffusion
+  - draft-model
+  - mlx
+---
+
+# Qwen3.8-27B DFlash2 generation-3 combined head
+
+This repository contains the promoted DFlash2 draft head for
+[`Qwen/Qwen3.8-27B`](https://huggingface.co/Qwen/Qwen3.8-27B). It is not a
+standalone language model and has no tokenizer: a compatible DFlash2 runtime
+must load it alongside the target verifier.
+
+The head descends from the public Qwen3.8-27B DFlash2 draft model. It was
+adapted with target-verified cycles from the incumbent legacy corpus and a
+session-grouped Fable agentic corpus. The final checkpoint combines a
+generation-2 Fable adaptation at strength `0.0625` with a generation-3
+combined-domain adaptation at strength `0.1875`.
+
+No trace dataset, prompt file, capture cycle, selector tensor, or W&B log is
+included in this repository. Only the final draft checkpoint, configuration,
+blend metadata, and evaluation summary are published.
+
+## Acceptance results
+
+All acceptance runs used greedy target-controlled generation on one Apple M4
+Max through DFlash MLX. The target was the calibrated Qwen3.8-27B MXFP4
+runtime. `Draft quantization` was disabled for this head.
+
+| Paired gate | Incumbent | This head | Mean delta | Wins / ties / losses |
+| --- | ---: | ---: | ---: | ---: |
+| Legacy 24-prompt corpus | `0.6712239583` | `0.6718750000` | `+0.0006510417` | 5 / 14 / 5 |
+| Fable 25-prompt held-out | `0.6471875000` | `0.6492187500` | `+0.0020312500` | 11 / 5 / 9 |
+
+The Fable live-gate prompts were excluded from the generation-2 and combined
+training splits. The legacy corpus had already been used to train the
+incumbent, so it remained in the combined training capture for a paired
+comparison.
+
+## Exact parity
+
+The exact 256-token baseline-versus-DFlash smoke gate passed:
+
+- `token_match`: `true`
+- smoke acceptance: `0.7109375`
+- diagnostic baseline throughput: `21.9615` tok/s
+- diagnostic DFlash throughput: `33.5111` tok/s
+
+Throughput is diagnostic only. Thermal pressure and co-tenant machine load
+were not controlled.
+
+## Training summary
+
+- combined capture: 11,273 target-verified cycles
+- prompt groups: 127
+- legacy contribution: 1,981 cycles from 24 prompts
+- Fable contribution: 9,292 cycles from 103 training prompts
+- excluded Fable live gate: 25 prompts
+- teacher signal: target top-64 logits and greedy target tokens
+- final full-run held-out prefix accuracy: `0.7932212528`
+- final full-run held-out top-1 accuracy: `0.8714877986`
+
+The capture teacher was the calibrated MXFP4 Qwen3.8-27B target runtime, not
+the pinned BF16 Qwen3.8-27B checkpoint. This precision choice is part of the
+training provenance.
+
+## Checkpoint integrity
+
+- `model.safetensors` SHA-256:
+  `c90e0fc65dddbe4107b1b417d771c3be188c87197e89cf4b14415f56abc866f0`
+- `config.json` SHA-256:
+  `873e3556509b0da06e29654ba00d4944888d4b5e8a33afde25f7eb27d321e980`
+- `blend.json` SHA-256:
+  `e53a6bb5e890d7e3368053302189766c3ef68d2eb7887152ece1ea9eac4f64d5`
+
+## Intended use and limits
+
+This artifact is for lossless speculative decoding research with a compatible
+DFlash2 implementation. Acceptance metrics do not establish broad task quality,
+natural-EOS behavior, multilingual performance, long-context performance, or a
+controlled thermal throughput win. The Fable held-out prompts come from the
+same eval shard as the Fable training prompts, so that gate does not prove
+out-of-domain generalization.

blend.json

@@ -0,0 +1 @@
+{"g2_strength": 0.0625, "g3_strength": 0.1875}

config.json

@@ -0,0 +1,57 @@
+{
+  "architectures": [
+    "DFlash2DraftModel"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": null,
+  "is_causal": false,
+  "dflash_config": {
+    "block_size": 8,
+    "conv_group_size": 16,
+    "conv_kernel_size": 2,
+    "mask_token_id": 248070,
+    "selector_rank": 256,
+    "selector_top_k": 16,
+    "target_layer_ids": [
+      5,
+      19,
+      33,
+      47,
+      61
+    ]
+  },
+  "dtype": "bfloat16",
+  "eos_token_id": 248044,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 17408,
+  "layer_types": [
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention"
+  ],
+  "max_position_embeddings": 262144,
+  "max_window_layers": 5,
+  "model_type": "qwen3",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 5,
+  "num_key_value_heads": 8,
+  "num_target_layers": 64,
+  "pad_token_id": 248044,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 10000000,
+    "rope_type": "default"
+  },
+  "sliding_window": 2048,
+  "tie_word_embeddings": false,
+  "transformers_version": "5.15.0",
+  "use_cache": true,
+  "use_sliding_window": true,
+  "vocab_size": 248320
+}

metrics.json

@@ -0,0 +1,65 @@
+{
+  "schema_version": 1,
+  "decision": "promoted",
+  "model_sha256": "c90e0fc65dddbe4107b1b417d771c3be188c87197e89cf4b14415f56abc866f0",
+  "config_sha256": "873e3556509b0da06e29654ba00d4944888d4b5e8a33afde25f7eb27d321e980",
+  "blend_sha256": "e53a6bb5e890d7e3368053302189766c3ef68d2eb7887152ece1ea9eac4f64d5",
+  "blend": {
+    "generation_2_strength": 0.0625,
+    "generation_3_strength": 0.1875
+  },
+  "runtime": {
+    "implementation": "DFlash MLX",
+    "revision": "dcf93d3d07aadc5a2e74fab7e44c63209d9f37a3",
+    "hardware": "Apple M4 Max",
+    "target": "calibrated Qwen3.8-27B MXFP4 runtime",
+    "draft_quantization": "none"
+  },
+  "exact_parity": {
+    "token_match": true,
+    "acceptance_ratio": 0.7109375,
+    "baseline_tps_diagnostic": 21.961506855085254,
+    "dflash_tps_diagnostic": 33.51110051912492
+  },
+  "paired_acceptance": [
+    {
+      "gate": "legacy24",
+      "control_mean_acceptance": 0.6712239583333334,
+      "candidate_mean_acceptance": 0.671875,
+      "mean_delta": 0.0006510416666666297,
+      "candidate_wins": 5,
+      "ties": 14,
+      "losses": 5
+    },
+    {
+      "gate": "fable-heldout25",
+      "control_mean_acceptance": 0.6471875,
+      "candidate_mean_acceptance": 0.64921875,
+      "mean_delta": 0.002031249999999929,
+      "candidate_wins": 11,
+      "ties": 5,
+      "losses": 9
+    }
+  ],
+  "training_capture": {
+    "cycle_count": 11273,
+    "prompt_group_count": 127,
+    "legacy_cycles": 1981,
+    "legacy_prompts": 24,
+    "fable_training_cycles": 9292,
+    "fable_training_prompts": 103,
+    "fable_heldout_prompts_excluded": 25,
+    "teacher_target_topk": 64,
+    "teacher_precision": "calibrated MXFP4 target runtime"
+  },
+  "training_metrics": {
+    "best_validation_prefix_accuracy": 0.793221252777784,
+    "best_validation_top1_accuracy": 0.871487798557914
+  },
+  "non_claims": [
+    "Acceptance metrics do not prove broad task quality or natural-EOS behavior.",
+    "Throughput and memory are diagnostic because thermal pressure was not controlled.",
+    "The Fable held-out prompts come from the same eval shard as Fable training prompts.",
+    "No trace data or capture cycles are included in this model repository."
+  ]
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c90e0fc65dddbe4107b1b417d771c3be188c87197e89cf4b14415f56abc866f0
+size 3848817885