FMHA pads N to next multiple of 128. For N<<128 (like 5 tokens),
the 123 padded zero-K entries contribute exp(0)=1 to the softmax
denominator, diluting real attention weights by ~128/5 = 25.6x.
This caused the model to produce incoherent output for short prompts.
Fix: use SDPA for seq_len < 120 (no padding), FMHA for longer
sequences where the padding effect is negligible.
Also: SDPA path includes attention sinks (paper D5c), FMHA path
uses analytic sink correction via LSE.
Instead of SDPA with virtual sink position, use the production FMHA
kernel and apply the sink bias as a post-hoc correction on the output.
The correction is: O_sink = O_raw * exp(lse) / (exp(lse) + exp(sink))
This simulates the attention sink (paper D5c) without modifying the
FMHA kernel. The sink absorbs some attention mass, reducing the
normalization constant and scaling down the output.
E2M1 magnitudes are [0, 0.5, 1, 1.5, 2, 3, 4, 6] NOT [0, 2, 3, 4, 6, 8, 12, 24].
The old LUT was 4x the correct values, causing every NVFP4 dequantized
weight to be 4x too large. This compounded across 61 layers, causing
the residual stream to explode and producing gibberish output.
This is the root cause of the residual growth and incoherent generation.
Replace custom mHCBlock with wrapper around the tested production
mHCLayer class. This eliminates any bugs in my custom implementation
and uses the same code path that the model was designed for.
Weight mapping: fn[0:4]=W_pre, fn[4:8]=W_post, fn[8:24]=W_res
base[0:4]=S_pre, base[4:8]=S_post, base[8:24]=S_res
scale[0]=alpha_pre, scale[1]=alpha_post, scale[2]=alpha_res
Checkpoint order is [pre, post, res] not [pre, res, post]:
- base[0:4] = S_pre, base[4:8] = S_post, base[8:24] = S_res
- scale[0] = alpha_pre, scale[1] = alpha_post, scale[2] = alpha_res
- W_stacked rows: [W_pre(4), W_post(4), W_res(16)]
- Projection split: A_raw=proj[:,0:4], C_raw=proj[:,4:8], B_raw=proj[:,8:24]
This was causing B_l to be near-identity and C_l to be near-2.0,
leading to exponential residual stream growth.
MoE:
- Hash routing (first 3 layers): tid2eid lookup → 6 experts, uniform weights
- Dense routing (remaining): sqrt(softplus(gate)) → top-6 → renormalize
- 384 NVFP4 experts, each gate+up+down with SiGLU clamping
- Weighted combine × routed_scaling_factor + shared expert
KV cache:
- SimpleKVCache: BF16 flat (1, max_seq, hd) per layer
- Appends new K,V each decode step
- FMHA now attends over full cached sequence (not just current token)
- RoPE applied per-position on K cache
This should produce meaningful output — the model now has all
architectural components except proper mHC normalization.
Without routed experts, F_out is always positive, causing unbounded
growth. Emergency RMSNorm on the residual keeps values bounded.
Remove once MoE is wired.
