Portable Colab/Kaggle diffusion lm-eval with SliceGPT and task workers

README.md

@@ -39,3 +39,11 @@ and both families, including Dream's pinned HF generation code. It checks reload
 logits, actual harness likelihood/generation, CPU/disk dispatch, and CUDA FP16
 with CPU offload when a local GPU is available. These checks do not replace a
 full 7B/8B benchmark run on the target platform.
+
+
+The notebook displays its effective config and exact adapter parameters before
+loading weights. To see the same preview from a terminal, add --show-config to
+your evaluation command. Denoising steps come from llada_steps/dream_steps
+(256); llada_generation_* fields are standalone helper defaults. Change
+MC_BATCH_SIZE, DTYPE, or RUN_PPL in the notebook settings cell to control its
+explicit overrides. RUN_PPL=True includes the shared perplexity settings.

bundle_manifest.json

@@ -1,6 +1,6 @@
 {
   "run_id": "run-9fc515e82f313fab",
-  "zip_sha256": "b411146207493dffc19ab59daf94dc7738d349a1d88bb7e0c4c4b6f546c9e6ec",
+  "zip_sha256": "e0860d29f84d16a6f75a934b3ff8e1f1b0185645b3ff677a550907177c3fc5d2",
   "available_checkpoints": 27,
   "shared_config": "Pruning-LLMs/diffusion_eval_and_vis/config.py",
   "smoke_tests": "All three methods on tiny LLaDA/Dream: likelihood, generation, CPU/disk and CUDA fp16 offload, reload agreement; task sharding and merging"

portable_eval.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b411146207493dffc19ab59daf94dc7738d349a1d88bb7e0c4c4b6f546c9e6ec
-size 37700
+oid sha256:e0860d29f84d16a6f75a934b3ff8e1f1b0185645b3ff677a550907177c3fc5d2
+size 38470

portable_lm_eval.ipynb

@@ -51,40 +51,65 @@
     },
     {
       "cell_type": "code",
-      "execution_count": null,
       "metadata": {},
-      "outputs": [],
       "source": [
         "REPO = 'vanshnawander/structured-llada-8b-instruct-sparsegpt-20-9fc515e82f313fab'\n",
-        "# Choose a different checkpoint in each session, or split tasks with these settings.\n",
         "WORKER_INDEX = 0\n",
         "NUM_WORKERS = 1\n",
-        "# MC remains 128. Small GPUs use mini-batches of 1 and automatic CPU/disk offload.\n",
-        "# dtype=auto uses bf16 when supported and float16 otherwise; actual precision is recorded.\n",
+        "\n",
+        "# Only these platform choices override the shared evaluation config.\n",
+        "MC_BATCH_SIZE = 1     # Shared default is 32; mc_num stays 128.\n",
+        "DTYPE = 'auto'        # Shared default is bfloat16; auto supports FP16-only GPUs.\n",
+        "RUN_PPL = False       # Set True to include the configured WikiText-2/C4 perplexities.\n",
+        "\n",
         "COMMON = [sys.executable, str(EVAL_ROOT / 'runner.py'), '--repo', REPO,\n",
-        "          '--batch-size', '1', '--dtype', 'auto', '--device-map', 'auto',\n",
+        "          '--batch-size', str(MC_BATCH_SIZE), '--dtype', DTYPE, '--device-map', 'auto',\n",
         "          '--gpu-memory-ratio', '0.65', '--output', 'results']\n",
-        "subprocess.run(COMMON + ['--smoke'], check=True)\n"
+        "FULL = COMMON.copy()\n",
+        "FULL[FULL.index('--output') + 1] = 'full_results'\n",
+        "FULL += ['--worker-index', str(WORKER_INDEX), '--num-workers', str(NUM_WORKERS)]\n",
+        "if not RUN_PPL:\n",
+        "    FULL.append('--skip-ppl')\n"
       ],
+      "execution_count": null,
+      "outputs": [],
       "id": "cell-4"
     },
     {
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "Run a short harness check first. This result is a smoke test, not the full benchmark. Re-running skips matching completed tasks. Full benchmark output uses a separate directory."
+        "Review the full benchmark settings before loading a checkpoint. This prints every effective config value and the exact parameters passed to the chosen adapter. Both families use 256 generation steps by default; LLaDA uses generation/block length 1024 and Dream generation length 256. MC likelihood scoring uses mc_num=128. The separate llada_generation_* helper defaults do not control the benchmark.\n"
       ],
       "id": "cell-5"
     },
     {
       "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "subprocess.run(FULL + ['--show-config'], check=True)\n"
+      ],
       "execution_count": null,
+      "outputs": [],
+      "id": "cell-6"
+    },
+    {
+      "cell_type": "code",
       "metadata": {},
+      "source": [
+        "subprocess.run(COMMON + ['--smoke'], check=True)\n"
+      ],
+      "execution_count": null,
       "outputs": [],
+      "id": "cell-7"
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
       "source": [
-        "subprocess.run(COMMON + ['--tasks', 'piqa', '--limit', '2', '--skip-ppl'], check=True)\n"
+        "Run a short harness check first. This result is a smoke test, not the full benchmark. Re-running skips matching completed tasks. Full benchmark output uses a separate directory."
       ],
-      "id": "cell-6"
+      "id": "cell-8"
     },
     {
       "cell_type": "code",
@@ -92,20 +117,27 @@
       "metadata": {},
       "outputs": [],
       "source": [
-        "FULL = COMMON.copy()\n",
-        "FULL[FULL.index('--output') + 1] = 'full_results'\n",
-        "subprocess.run(FULL + ['--skip-ppl', '--worker-index', str(WORKER_INDEX),\n",
-        "                       '--num-workers', str(NUM_WORKERS)], check=True)\n"
+        "subprocess.run(COMMON + ['--tasks', 'piqa', '--limit', '2', '--skip-ppl'], check=True)\n"
       ],
-      "id": "cell-7"
+      "id": "cell-9"
+    },
+    {
+      "cell_type": "code",
+      "metadata": {},
+      "source": [
+        "subprocess.run(FULL, check=True)\n"
+      ],
+      "execution_count": null,
+      "outputs": [],
+      "id": "cell-10"
     },
     {
       "cell_type": "markdown",
       "metadata": {},
       "source": [
-        "The full run uses all seven configured tasks unless you provide --tasks. GSM8K keeps the harness's built-in five shots; other tasks use zero shots. Remove --skip-ppl to compute configured WikiText-2/C4 perplexities too. For parallel task evaluation, use NUM_WORKERS=2 with WORKER_INDEX=0 in one session and 1 in the other, using the same checkpoint and settings."
+        "The full run uses all seven configured tasks unless you provide --tasks. GSM8K keeps the harness's built-in five shots; other tasks use zero shots. Set RUN_PPL=True in the settings cell to compute configured WikiText-2/C4 perplexities too, then rerun the preview. For parallel task evaluation, use NUM_WORKERS=2 with WORKER_INDEX=0 in one session and 1 in the other, using the same checkpoint and settings.\n"
       ],
-      "id": "cell-8"
+      "id": "cell-11"
     },
     {
       "cell_type": "code",
@@ -121,7 +153,7 @@
         "except ImportError:\n",
         "    print('Download evaluation_results.zip from the notebook output/files panel.')\n"
       ],
-      "id": "cell-9"
+      "id": "cell-12"
     },
     {
       "cell_type": "markdown",
@@ -129,7 +161,7 @@
       "source": [
         "Keep the result ZIP before the session expires. Extract ZIPs from multiple workers into one results directory, then run runner.py --merge that_directory --output merged_results.json. It groups results by checkpoint revision and evaluation settings and rejects conflicting duplicate task scores."
       ],
-      "id": "cell-10"
+      "id": "cell-13"
     }
   ],
   "metadata": {

runner.py

@@ -196,22 +196,29 @@ def load_checkpoint(path, cfg, device_map="auto", gpu_memory_ratio=0.65, cpu_gib
     return model, tokenizer
 
 
+def adapter_parameters(cfg, family):
+    """The exact config values passed to the benchmark adapter."""
+    common = {"batch_size": cfg.batch_size, "mc_num": cfg.mc_num, "device": cfg.device}
+    if family == "dream":
+        return {**common, "add_bos_token": cfg.add_bos_token and not cfg.no_add_bos_token,
+                "max_length": cfg.max_length, "nll_type": cfg.nll_type, "log_type": cfg.log_type,
+                "sampling_eps": cfg.sampling_eps, "classifier_free_guidance": cfg.dream_cfg,
+                "diffusion_steps": cfg.dream_steps, "gen_length": cfg.dream_gen_length,
+                "gen_alg": cfg.dream_gen_alg, "gen_temperature": cfg.dream_gen_temperature}
+    return {**common, "mask_id": cfg.llada_mask_id, "dtype": cfg.dtype,
+            "is_check_greedy": cfg.is_check_greedy, "cfg": cfg.llada_cfg,
+            "steps": cfg.llada_steps, "gen_length": cfg.llada_gen_length,
+            "block_length": cfg.llada_block_length, "remasking": cfg.llada_remasking}
+
+
 def make_adapter(model, tokenizer, cfg, family, checkpoint=""):
+    parameters = adapter_parameters(cfg, family)
     if family == "dream":
         from diffusion_eval_and_vis.eval_dream import DreamLM
-        return DreamLM(model, tokenizer, batch_size=cfg.batch_size, mc_num=cfg.mc_num,
-                       device=cfg.device, add_bos_token=cfg.add_bos_token and not cfg.no_add_bos_token,
-                       max_length=cfg.max_length, nll_type=cfg.nll_type, log_type=cfg.log_type,
-                       sampling_eps=cfg.sampling_eps, classifier_free_guidance=cfg.dream_cfg,
-                       diffusion_steps=cfg.dream_steps, gen_length=cfg.dream_gen_length,
-                       gen_alg=cfg.dream_gen_alg, gen_temperature=cfg.dream_gen_temperature)
+        return DreamLM(model, tokenizer, **parameters)
     from diffusion_eval_and_vis.eval_diffusion import LLaDiffusionLM
     return LLaDiffusionLM(model_path=checkpoint, model_instance=model, tokenizer_instance=tokenizer,
-                         mask_id=cfg.llada_mask_id, dtype=cfg.dtype, device=cfg.device,
-                         batch_size=cfg.batch_size, mc_num=cfg.mc_num,
-                         is_check_greedy=cfg.is_check_greedy, cfg=cfg.llada_cfg,
-                         steps=cfg.llada_steps, gen_length=cfg.llada_gen_length,
-                         block_length=cfg.llada_block_length, remasking=cfg.llada_remasking)
+                         **parameters)
 
 
 def merge_results(root, destination):
@@ -241,6 +248,7 @@ def main():
     parser.add_argument("--checkpoint", help="Already downloaded local checkpoint")
     parser.add_argument("--family", choices=["llada", "dream"])
     parser.add_argument("--list", action="store_true")
+    parser.add_argument("--show-config", action="store_true", help="Print effective config and adapter parameters without loading weights")
     parser.add_argument("--owner", default=DEFAULT_OWNER)
     parser.add_argument("--run", default=DEFAULT_RUN)
     parser.add_argument("--tasks", nargs="+")
@@ -267,7 +275,7 @@ def main():
     if args.merge:
         print({"merged_evaluations": merge_results(args.merge, args.output)})
         return
-    if not args.repo and not args.checkpoint:
+    if not args.repo and not args.checkpoint and not args.show_config:
         parser.error("Supply --repo or --checkpoint")
     import torch
     from huggingface_hub import HfApi, snapshot_download
@@ -286,6 +294,15 @@ def main():
         raise RuntimeError("Enable a GPU runtime or pass --device cpu")
     if cfg.dtype == "bfloat16" and cfg.device.startswith("cuda") and not torch.cuda.is_bf16_supported():
         raise RuntimeError("This GPU needs --dtype auto or --dtype float16; the override is recorded in results")
+    if args.show_config:
+        family = args.family or ("dream" if "structured-dream-" in (args.repo or "") else "llada")
+        tasks = args.tasks or (cfg.dream_tasks or cfg.llada_tasks if family == "dream" else cfg.llada_tasks)
+        print(json.dumps({"family": family, "eval_config": asdict(cfg),
+                          "adapter_parameters": adapter_parameters(cfg, family),
+                          "worker_tasks": split_tasks(tasks, args.worker_index, args.num_workers),
+                          "perplexity_enabled": not cfg.skip_ppl,
+                          "note": "llada_generation_* are separate helper defaults; the benchmark uses llada_steps/llada_gen_length/llada_block_length."}, indent=2))
+        return
     token = os.environ.get("HF_TOKEN", False)
     if args.checkpoint:
         checkpoint, revision = args.checkpoint, args.revision or "local"