katostrofik/qwen35moe-blackhole-p100-slim-gate-up-fused-2026-05-24

Qwen35MoE Blackhole P100 Slim Gate-Up Fused Artifact

This artifact is the large-file companion to the GitHub handoff repo/branch for the single-card Qwen 3.5/Qwen3.5-MoE-family Blackhole P100 bring-up.

It contains the TTNN tensorbin payload needed by:

python llm/qwen35moe_resident_cache_load_walk.py \
  --bundle llm/cache/qwen36_resident_cache/full_resident_p100_slim_gate_up_fused_bundle.json \
  --streaming

python llm/qwen35moe_showable_baseline.py \
  --bundle llm/cache/qwen36_resident_cache/full_resident_p100_slim_gate_up_fused_bundle.json \
  --tokenizer-json llm/models/Qwen3.6-35B-A3B-FP8/tokenizer.json

Important Claim

This was validated locally on one Blackhole P150A. The slim resident payload and runtime estimate should fit a 28 GiB P100, but direct P100 validation is still pending.

Contents

Required model/cache payload:

• —llm/cache/qwen36_resident_cache/full_resident_p100_slim_gate_up_fused_bundle.json
• —llm/cache/qwen36_resident_cache/full_resident_p100_slim_gate_up_fused_report.json
• —llm/cache/qwen36_resident_cache/materialized_mixed_fit_dense_all
• —llm/cache/qwen36_resident_cache/materialized_mixed_fit_global
• —llm/cache/qwen36_resident_cache/materialized_mixed_fit_layer0_routed_banks
• —llm/cache/qwen36_resident_cache/materialized_mixed_fit_routed_banks_layers1_39
• —llm/cache/qwen36_resident_cache/materialized_mixed_fit_routed_gate_up_fused
• —llm/cache/qwen36_resident_cache/materialized_mixed_fit_routers_all
• —llm/cache/qwen36_resident_cache/materialized_mixed_fit_shared_experts_all
• —llm/cache/qwen36_resident_cache/global_weights_bf16_embedding
• —llm/models/Qwen3.6-35B-A3B-FP8/tokenizer.json

Reference outputs:

• —llm/cache/qwen36_resident_cache/full_resident_p100_slim_gate_up_fused_load_walk_streaming.json
• —llm/cache/qwen36_resident_cache/full_resident_p100_slim_gate_up_fused_load_walk_resident.json
• —llm/cache/qwen36_resident_cache/qwen35moe_p100_slim_showable_baseline.json

Size

• —Slim resident bundle tensorbins: about 18.81 GiB
• —BF16 embedding tensor: about 0.95 GiB
• —Tokenizer JSON: about 12 MB
• —Estimated loaded runtime, including embedding and decode state: about 19.85 GiB

Expected Baseline Output

Local AI empowers users to run machine learning models directly on their own devices, ensuring data privacy and eliminating reliance on external cloud servers.<|im_end|>