qnn/sm8750 4k: HuggingFaceTB/SmolLM2-135M-Instruct@12fd25f77366

.gitattributes

@@ -38,3 +38,4 @@ xnnpack/SmolLM2-135M-Instruct-8da4w-2k.pte filter=lfs diff=lfs merge=lfs -text
 xnnpack/SmolLM2-135M-Instruct-8da4w-8k.pte filter=lfs diff=lfs merge=lfs -text
 xnnpack/SmolLM2-135M-Instruct-8da4w-16k.pte filter=lfs diff=lfs merge=lfs -text
 qnn/sm8750/SmolLM2-135M-Instruct-qnn-hybrid-2k.pte filter=lfs diff=lfs merge=lfs -text
+qnn/sm8750/SmolLM2-135M-Instruct-qnn-hybrid-4k.pte filter=lfs diff=lfs merge=lfs -text

README.md

@@ -29,6 +29,7 @@ Every backend is exported at every context window the runner could build (2k to
 | XNNPACK (CPU) | any arm64 | [`xnnpack/SmolLM2-135M-Instruct-8da4w-8k.pte`](xnnpack/SmolLM2-135M-Instruct-8da4w-8k.pte) | 8,192 tokens | 0.11 GB | passed ("Paris") |
 | XNNPACK (CPU) | any arm64 | [`xnnpack/SmolLM2-135M-Instruct-8da4w-16k.pte`](xnnpack/SmolLM2-135M-Instruct-8da4w-16k.pte) | 16,384 tokens | 0.11 GB | passed ("Paris") |
 | Qualcomm QNN (HTP) | SM8750 (Snapdragon 8 Elite) | [`qnn/sm8750/SmolLM2-135M-Instruct-qnn-hybrid-2k.pte`](qnn/sm8750/SmolLM2-135M-Instruct-qnn-hybrid-2k.pte) | 2,048 tokens | 0.20 GB | structure checked (no host NPU runtime) |
+| Qualcomm QNN (HTP) | SM8750 (Snapdragon 8 Elite) | [`qnn/sm8750/SmolLM2-135M-Instruct-qnn-hybrid-4k.pte`](qnn/sm8750/SmolLM2-135M-Instruct-qnn-hybrid-4k.pte) | 4,096 tokens | 0.21 GB | structure checked (no host NPU runtime) |
 
 Tokenizer: [`tokenizer.json`](tokenizer.json), copied unchanged from the source repo. Each backend folder has a `config.json` listing every window as a variant with the metadata the `.pte` reports, and an `export-report-<window>.json` per file with the full export record.
 

qnn/sm8750/SmolLM2-135M-Instruct-qnn-hybrid-4k.pte

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15aa9bb8fb5bc2807fceb0cbb71b2296ce96933f3f708fa02acbb91344012701
+size 205647488

qnn/sm8750/config.json

@@ -22,6 +22,22 @@
         "get_vocab_size": 49152,
         "get_n_layers": 30
       }
+    },
+    {
+      "file": "SmolLM2-135M-Instruct-qnn-hybrid-4k.pte",
+      "size_bytes": 205647488,
+      "sha256": "15aa9bb8fb5bc2807fceb0cbb71b2296ce96933f3f708fa02acbb91344012701",
+      "context": 4096,
+      "source_revision": "12fd25f77366fa6b3b4b768ec3050bf629380bac",
+      "quantization": "QNN HTP, ExecuTorch smollm2_135m recipe",
+      "fits_phone_budget": null,
+      "methods": {
+        "get_max_context_len": 4096,
+        "get_max_seq_len": 4096,
+        "get_bos_id": 1,
+        "get_vocab_size": 49152,
+        "get_n_layers": 30
+      }
     }
   ],
   "qnn_sdk_version": "2.37.0.250724"

qnn/sm8750/export-report-4k.json

@@ -0,0 +1,691 @@
+{
+  "backend": "qnn",
+  "target": "sm8750",
+  "target_name": "Snapdragon 8 Elite",
+  "tokenizer": "tokenizer.json",
+  "output_repo": "experimentalmachines/SmolLM2-135M-Instruct-ExecuTorch",
+  "source": {
+    "id": "HuggingFaceTB/SmolLM2-135M-Instruct",
+    "sha": "12fd25f77366fa6b3b4b768ec3050bf629380bac",
+    "created_at": "2024-10-31T13:41:10+00:00",
+    "total_params": 134515008,
+    "family": "llama",
+    "variant": "instruct",
+    "license": {
+      "license": "apache-2.0"
+    },
+    "license_files": []
+  },
+  "toolchain": {
+    "executorch": "1.4.0",
+    "torch": "2.13.0+cpu",
+    "torchao": "0.18.0",
+    "transformers": "5.0.0rc1",
+    "lm_eval": "0.4.5",
+    "qairt": "2.37.0.250724"
+  },
+  "recipe": {
+    "decoder_model": "smollm2_135m",
+    "params": "executorch: examples/models/smollm2/135M_config.json",
+    "model_mode": "hybrid",
+    "prefill_ar_len": 128,
+    "max_context_len": 4096,
+    "calibration": {
+      "tasks": [
+        "wikitext"
+      ],
+      "limit": 1
+    },
+    "label": "QNN HTP, ExecuTorch smollm2_135m recipe",
+    "description": "ExecuTorch 1.4.0 Qualcomm static LLM (`examples/qualcomm/oss_scripts/llama`, `--decoder_model smollm2_135m`): the quantization recipe ExecuTorch registers for this model, calibrated on wikitext (1 sample), hybrid prefill (128 tokens per step) and decode graphs compiled with QAIRT 2.37.0.250724 for SM8750 (Snapdragon 8 Elite)."
+  },
+  "window": {
+    "context": 4096,
+    "reason": "forced with --context (static NPU graphs)",
+    "kv_cache_bytes_per_token": null
+  },
+  "files": [
+    {
+      "path": "qnn/sm8750/SmolLM2-135M-Instruct-qnn-hybrid-4k.pte",
+      "bytes": 205647488,
+      "sha256": "15aa9bb8fb5bc2807fceb0cbb71b2296ce96933f3f708fa02acbb91344012701"
+    }
+  ],
+  "host": {
+    "nproc": 4,
+    "mem_total_bytes": 16766414848,
+    "swap_total_bytes": 25769799680,
+    "cpu_model": "AMD EPYC 7763 64-Core Processor",
+    "peak_rss_export_bytes": 10365153280,
+    "peak_swap_used_bytes": 24576,
+    "min_mem_available_bytes": 6959456256,
+    "peak_in_use_bytes": 9806979072,
+    "peak_in_use_at": "2026-09-13T15:28:17Z",
+    "export_seconds": 1814.8,
+    "total_seconds": 1820.2
+  },
+  "metadata": {
+    "get_ar_len": 1,
+    "get_bos_id": 1,
+    "get_dim": 576,
+    "get_eos_id": 2,
+    "get_head_dim": 64,
+    "get_kv_io_bit_width": 16,
+    "get_kv_output_0_quant_attr": [
+      1.5267929484252818e-05,
+      39862,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_10_quant_attr": [
+      0.00010680226841941476,
+      34307,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_11_quant_attr": [
+      0.0002643346379045397,
+      33586,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_12_quant_attr": [
+      0.00010360516171203926,
+      35927,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_13_quant_attr": [
+      0.0002805650874506682,
+      31995,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_14_quant_attr": [
+      0.00015581400657538325,
+      34356,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_15_quant_attr": [
+      0.00032592701609246433,
+      35482,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_16_quant_attr": [
+      0.0001547761494293809,
+      32278,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_17_quant_attr": [
+      0.000292566284770146,
+      32587,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_18_quant_attr": [
+      0.00015396272647194564,
+      32939,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_19_quant_attr": [
+      0.0003059740993194282,
+      31888,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_1_quant_attr": [
+      0.00030670216074213386,
+      30269,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_20_quant_attr": [
+      0.00013691841741092503,
+      29959,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_21_quant_attr": [
+      0.00028609068249352276,
+      33451,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_22_quant_attr": [
+      0.00015312536561395973,
+      35678,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_23_quant_attr": [
+      0.0005065652658231556,
+      33834,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_24_quant_attr": [
+      0.00012949334632139653,
+      35488,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_25_quant_attr": [
+      0.0003119675093330443,
+      32065,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_26_quant_attr": [
+      0.00010754493996500969,
+      28212,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_27_quant_attr": [
+      0.00033483997685834765,
+      34778,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_28_quant_attr": [
+      0.00016533619782421738,
+      28571,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_29_quant_attr": [
+      0.00032498195650987327,
+      32338,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_2_quant_attr": [
+      9.45554711506702e-05,
+      34333,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_30_quant_attr": [
+      0.00015253313176799566,
+      29696,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_31_quant_attr": [
+      0.0003206882975064218,
+      33783,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_32_quant_attr": [
+      0.0001389853423461318,
+      34639,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_33_quant_attr": [
+      0.0002721503551583737,
+      33244,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_34_quant_attr": [
+      0.00017622944142203778,
+      34816,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_35_quant_attr": [
+      0.0003714370250236243,
+      34989,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_36_quant_attr": [
+      0.00027163943741470575,
+      27939,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_37_quant_attr": [
+      0.00045417060027830303,
+      33588,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_38_quant_attr": [
+      0.00020489229063969105,
+      34867,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_39_quant_attr": [
+      0.0003060104791074991,
+      33456,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_3_quant_attr": [
+      0.0003054087865166366,
+      32608,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_40_quant_attr": [
+      0.0002207932120654732,
+      29549,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_41_quant_attr": [
+      0.0002991298970300704,
+      33455,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_42_quant_attr": [
+      0.00016994334873743355,
+      34993,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_43_quant_attr": [
+      0.00036768231075257063,
+      32620,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_44_quant_attr": [
+      0.0002000726235564798,
+      31367,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_45_quant_attr": [
+      0.00032547954469919205,
+      31660,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_46_quant_attr": [
+      0.00034692950430326164,
+      32081,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_47_quant_attr": [
+      0.0003443109162617475,
+      33475,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_48_quant_attr": [
+      0.000409788015531376,
+      32635,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_49_quant_attr": [
+      0.000334571028361097,
+      35048,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_4_quant_attr": [
+      0.00010261722491122782,
+      31880,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_50_quant_attr": [
+      0.00028442314942367375,
+      27922,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_51_quant_attr": [
+      0.00028256839141249657,
+      34140,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_52_quant_attr": [
+      0.0003381276910658926,
+      33625,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_53_quant_attr": [
+      0.00028591937734745443,
+      32368,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_54_quant_attr": [
+      0.00026667446945793927,
+      32573,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_55_quant_attr": [
+      0.00026856877957470715,
+      32578,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_56_quant_attr": [
+      0.0004002549685537815,
+      30346,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_57_quant_attr": [
+      0.00023681700986344367,
+      34943,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_58_quant_attr": [
+      0.00040069734677672386,
+      32320,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_59_quant_attr": [
+      0.0002558212145231664,
+      31997,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_5_quant_attr": [
+      0.00031697546364739537,
+      33672,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_6_quant_attr": [
+      0.0001312306267209351,
+      29326,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_7_quant_attr": [
+      0.0003144744841847569,
+      32406,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_8_quant_attr": [
+      0.00012001295544905588,
+      37138,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_kv_output_9_quant_attr": [
+      0.0002740690833888948,
+      33056,
+      0,
+      65535,
+      "torch.int32"
+    ],
+    "get_logits_scale": 0.0012609856203198433,
+    "get_logits_zero_point": 31912,
+    "get_max_batch_size": 1,
+    "get_max_context_len": 4096,
+    "get_max_seq_len": 4096,
+    "get_n_bos": 1,
+    "get_n_eos": 1,
+    "get_n_kv_heads": 3,
+    "get_n_layers": 30,
+    "get_use_kv_cache": true,
+    "get_vocab_size": 49152,
+    "methods": [
+      "get_ar_len",
+      "get_bos_id",
+      "get_dim",
+      "get_eos_id",
+      "get_head_dim",
+      "get_kv_io_bit_width",
+      "get_kv_output_0_quant_attr",
+      "get_kv_output_10_quant_attr",
+      "get_kv_output_11_quant_attr",
+      "get_kv_output_12_quant_attr",
+      "get_kv_output_13_quant_attr",
+      "get_kv_output_14_quant_attr",
+      "get_kv_output_15_quant_attr",
+      "get_kv_output_16_quant_attr",
+      "get_kv_output_17_quant_attr",
+      "get_kv_output_18_quant_attr",
+      "get_kv_output_19_quant_attr",
+      "get_kv_output_1_quant_attr",
+      "get_kv_output_20_quant_attr",
+      "get_kv_output_21_quant_attr",
+      "get_kv_output_22_quant_attr",
+      "get_kv_output_23_quant_attr",
+      "get_kv_output_24_quant_attr",
+      "get_kv_output_25_quant_attr",
+      "get_kv_output_26_quant_attr",
+      "get_kv_output_27_quant_attr",
+      "get_kv_output_28_quant_attr",
+      "get_kv_output_29_quant_attr",
+      "get_kv_output_2_quant_attr",
+      "get_kv_output_30_quant_attr",
+      "get_kv_output_31_quant_attr",
+      "get_kv_output_32_quant_attr",
+      "get_kv_output_33_quant_attr",
+      "get_kv_output_34_quant_attr",
+      "get_kv_output_35_quant_attr",
+      "get_kv_output_36_quant_attr",
+      "get_kv_output_37_quant_attr",
+      "get_kv_output_38_quant_attr",
+      "get_kv_output_39_quant_attr",
+      "get_kv_output_3_quant_attr",
+      "get_kv_output_40_quant_attr",
+      "get_kv_output_41_quant_attr",
+      "get_kv_output_42_quant_attr",
+      "get_kv_output_43_quant_attr",
+      "get_kv_output_44_quant_attr",
+      "get_kv_output_45_quant_attr",
+      "get_kv_output_46_quant_attr",
+      "get_kv_output_47_quant_attr",
+      "get_kv_output_48_quant_attr",
+      "get_kv_output_49_quant_attr",
+      "get_kv_output_4_quant_attr",
+      "get_kv_output_50_quant_attr",
+      "get_kv_output_51_quant_attr",
+      "get_kv_output_52_quant_attr",
+      "get_kv_output_53_quant_attr",
+      "get_kv_output_54_quant_attr",
+      "get_kv_output_55_quant_attr",
+      "get_kv_output_56_quant_attr",
+      "get_kv_output_57_quant_attr",
+      "get_kv_output_58_quant_attr",
+      "get_kv_output_59_quant_attr",
+      "get_kv_output_5_quant_attr",
+      "get_kv_output_6_quant_attr",
+      "get_kv_output_7_quant_attr",
+      "get_kv_output_8_quant_attr",
+      "get_kv_output_9_quant_attr",
+      "get_logits_scale",
+      "get_logits_zero_point",
+      "get_max_batch_size",
+      "get_max_context_len",
+      "get_max_seq_len",
+      "get_n_bos",
+      "get_n_eos",
+      "get_n_kv_heads",
+      "get_n_layers",
+      "get_use_kv_cache",
+      "get_vocab_size",
+      "kv_forward",
+      "prefill_forward"
+    ]
+  },
+  "smoke": {
+    "kind": "structural",
+    "passed": true,
+    "problems": [],
+    "methods": [
+      "get_ar_len",
+      "get_bos_id",
+      "get_dim",
+      "get_eos_id",
+      "get_head_dim",
+      "get_kv_io_bit_width",
+      "get_kv_output_0_quant_attr",
+      "get_kv_output_10_quant_attr",
+      "get_kv_output_11_quant_attr",
+      "get_kv_output_12_quant_attr",
+      "get_kv_output_13_quant_attr",
+      "get_kv_output_14_quant_attr",
+      "get_kv_output_15_quant_attr",
+      "get_kv_output_16_quant_attr",
+      "get_kv_output_17_quant_attr",
+      "get_kv_output_18_quant_attr",
+      "get_kv_output_19_quant_attr",
+      "get_kv_output_1_quant_attr",
+      "get_kv_output_20_quant_attr",
+      "get_kv_output_21_quant_attr",
+      "get_kv_output_22_quant_attr",
+      "get_kv_output_23_quant_attr",
+      "get_kv_output_24_quant_attr",
+      "get_kv_output_25_quant_attr",
+      "get_kv_output_26_quant_attr",
+      "get_kv_output_27_quant_attr",
+      "get_kv_output_28_quant_attr",
+      "get_kv_output_29_quant_attr",
+      "get_kv_output_2_quant_attr",
+      "get_kv_output_30_quant_attr",
+      "get_kv_output_31_quant_attr",
+      "get_kv_output_32_quant_attr",
+      "get_kv_output_33_quant_attr",
+      "get_kv_output_34_quant_attr",
+      "get_kv_output_35_quant_attr",
+      "get_kv_output_36_quant_attr",
+      "get_kv_output_37_quant_attr",
+      "get_kv_output_38_quant_attr",
+      "get_kv_output_39_quant_attr",
+      "get_kv_output_3_quant_attr",
+      "get_kv_output_40_quant_attr",
+      "get_kv_output_41_quant_attr",
+      "get_kv_output_42_quant_attr",
+      "get_kv_output_43_quant_attr",
+      "get_kv_output_44_quant_attr",
+      "get_kv_output_45_quant_attr",
+      "get_kv_output_46_quant_attr",
+      "get_kv_output_47_quant_attr",
+      "get_kv_output_48_quant_attr",
+      "get_kv_output_49_quant_attr",
+      "get_kv_output_4_quant_attr",
+      "get_kv_output_50_quant_attr",
+      "get_kv_output_51_quant_attr",
+      "get_kv_output_52_quant_attr",
+      "get_kv_output_53_quant_attr",
+      "get_kv_output_54_quant_attr",
+      "get_kv_output_55_quant_attr",
+      "get_kv_output_56_quant_attr",
+      "get_kv_output_57_quant_attr",
+      "get_kv_output_58_quant_attr",
+      "get_kv_output_59_quant_attr",
+      "get_kv_output_5_quant_attr",
+      "get_kv_output_6_quant_attr",
+      "get_kv_output_7_quant_attr",
+      "get_kv_output_8_quant_attr",
+      "get_kv_output_9_quant_attr",
+      "get_logits_scale",
+      "get_logits_zero_point",
+      "get_max_batch_size",
+      "get_max_context_len",
+      "get_max_seq_len",
+      "get_n_bos",
+      "get_n_eos",
+      "get_n_kv_heads",
+      "get_n_layers",
+      "get_use_kv_cache",
+      "get_vocab_size",
+      "kv_forward",
+      "prefill_forward"
+    ],
+    "delegates": {
+      "kv_forward": {
+        "backends": [
+          "QnnBackend"
+        ],
+        "delegate_calls": 1
+      },
+      "prefill_forward": {
+        "backends": [
+          "QnnBackend"
+        ],
+        "delegate_calls": 1
+      }
+    }
+  },
+  "run": {
+    "url": "https://github.com/ExperimentalMachines/executorch-model-exporter/actions/runs/34764080127",
+    "id": "34764080127"
+  }
+}