Update model lists

[bjarzemb]

Changes

1. Splitted priority lists per run configuration

Restructure each priority (p0, p1, p2) from a single flat model list into
per-run-configuration lists so that accuracy and performance, inference and
training can each carry their own model set:

p{0,1,2}/
  accuracy/
    inference/{torchbench,huggingface,timm}_models_list.txt
    training/{torchbench,huggingface,timm}_models_list.txt
  performance/
    inference/{torchbench,huggingface,timm}_models_list.txt
    training/{torchbench,huggingface,timm}_models_list.txt

2. Added new models to priority list

• huggingface (P0)
  • OPTForCausalLM (ALL)
  • XGLMForCausalLM (ALL)
  • M2M100ForConditionalGeneration (acc/inf, perf/inf)
• timm models (P2)
  • cait_m36_384 (ALL)
  • xcit_large_24_p8_224 (ALL)
• torchbench (P0)
  • torchrec_dlrm (ALL)
  • demucs (ALL)
  • detectron2_maskrcnn_r_50_c4 (ALL)
  • llava (ALL)
  • hf_BigBird (ALL)
  • hf_Longformer (ALL)
  • detectron2_fcos_r_50_fpn (acc/inf, perf/inf)
  • llama_v2_7b_16h (acc/inf, perf/inf)
  • moondream (acc/inf, perf/inf)
  • cm3leon_generate (perf/inf)

3. Removed models from priority list (because they are explicitly skipped in benchmarks/dynamo/torchbench.yaml:

• stable_diffusion_text_encoder
• stable_diffusion_unet
• doctr_det_predictor (acc/train, perf/train)
• doctr_reco_predictor (acc/train, perf/train)
• maml (acc/train, perf/train)
• pyhpc_equation_of_state (acc/train, perf/train)
• pyhpc_isoneutral_mixing (acc/train, perf/train)
• pyhpc_turbulent_kinetic_energy (acc/train, perf/train)
• sam_fast (acc/train, perf/train)

4. Removed duplicate HuggingFace entries from top-level list

• mistralai/Mistral-7B-Instruct-v0.3
• openai/gpt-oss-20b

[github-actions]

Performance outliers, please check!

• 🟡 [80%, 90%), may be fluctuations

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
huggingface_bfloat16_training	BlenderbotForCausalLM	0.869883	0.880032
huggingface_bfloat16_training	DistilBertForMaskedLM	0.909026	0.883223

[github-actions]

Performance outliers, please check!

• 🟡 [80%, 90%), may be fluctuations

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
huggingface_bfloat16_training	DistilBertForMaskedLM	0.883406	0.894316
huggingface_bfloat16_training	BlenderbotForCausalLM	0.890704	0.905839