Critical bug: checkpoint weights are (N_packed, K_packed) N-major format,
but make_b_k_major expects (E, K_packed, N_packed) input. Without the
permute, the K and N dimensions are swapped, producing garbage output
with wrong dimensions (e.g., q_a output was 3584 instead of 1536).
Also fix scale assembly: checkpoint scales are (N, K_sf) which should
use assemble_raw_scales_2d3d_3d_side (no transpose), not
assemble_scales_3d_side (which incorrectly transposes K_sf↔N).
Critical bug fix: weight_scale_2 (the second-level NVFP4 scale) was
being dropped entirely in the production pipeline. The dequant formula
is lut[w] * weight_scale * weight_scale_2, so weight_scale_2 must be
folded into the GEMM's global_scale_b parameter.
Fixes in:
- Nvfp4Linear: ws2 field, folded in finalize_weights()
- Nvfp4MoE: l1_ws2/l2_ws2 lists, folded in _ensure_stacked()
- Nvfp4SharedExpert: l1_ws2/l2_ws2 lists, folded in finalize_weights()
- single_shot_inference.py: pass weight_scale_2 through all loading paths
- Also fix missing o_a_prod key fallback in attention output
Nvfp4Linear causing CUDA context corruption (likely CuTeDSL JIT
triggered by _ensure_initialized). Disable for now to validate
the critical paths first:
- Production FMHA with sink bias
- Production MoE (Nvfp4MoE + Nvfp4SharedExpert)
- Production Router (dense/hash)
- Production mHC
Attention projections use reference dequant+matmul for now.
Will re-enable Nvfp4Linear after validating MoE path.
Build stacked (E, N, K) tensors incrementally on CPU, then move to GPU
in one shot. Avoids holding 384 individual expert weight+scale tensors
on GPU simultaneously (~3x memory savings per layer).
ROOT CAUSE: fmha_multitile_op.py padded N to 128 for TMA alignment
but then passed the PADDED N to the kernel as s_k (logical KV length).
This told the kernel all 128 entries were valid, so softmax ran over
zeros, diluting the result (e.g. 1 valid entry → softmax weight 1/128).
FIX: Pass N_orig (true sequence length) as s_k for softmax masking,
and N_padded (physical size) only for TMA descriptor creation.
The kernel's existing col < kv_len guard correctly excludes padded
entries from row_max and exp_sum calculations.
Files changed:
- fmha_multitile_capi.cu: accept N_orig + N_padded, use N_orig for
params.s_k and N_padded for TMA descriptors
- fmha_multitile_op.py: pass N_orig and N_padded separately
- single_shot_inference.py: removed SDPA fallback (kernel now correct)
input_scale is the activation quantization scale (for FP8 inputs).
Since we use BF16 activations, the weight dequant is simply:
lut[weight] * weight_scale * weight_scale_2
Folding input_scale in produced weights ~4000x too small,
causing all attention and FFN outputs to be effectively zero.
The model uses DeepseekV4HyperHead to project from the 4-stream mHC
residual to the final hidden state. Just taking stream 0 (X[:,0,:])
is WRONG — the hc_head learns how to combine the 4 streams.
Also:
- Remove --no-thinking mode (this is a reasoning model, it MUST think)
- Increase default max_tokens from 512 to 4096
- Load hc_head weights (fn, base, scale) from checkpoint
Bugs fixed (verified against HuggingFace DeepseekV4HyperConnection):
1. fn/base/scale ordering was [pre,comb,post], should be [pre,post,comb]
- Was applying Sinkhorn to post values and 2*sigmoid to comb values
- This caused residual to grow unbounded (no doubly-stochastic constraint)
2. comb (B_l) must be TRANSPOSED in post_block
- HF: comb.transpose(-1,-2) @ hidden_streams
- Was using B_l @ X_l without transpose
3. Sinkhorn must start from softmax(logits) + eps, not exp(logits)
- HF: softmax → col norm → (iters-1) alternating
- Was using exp → alternating (different convergence behavior)
4. Missing hc_eps on pre (A_l)
- HF: sigmoid(...) + hc_eps
- Was missing the eps guard
5. Renamed W_res→W_comb, S_res→S_comb, alpha_res→alpha_comb throughout
- Matches checkpoint naming and HF model
6. Fixed fallback mHC initialization to use new API
The HuggingFace reference treats attention sinks as a logit bias:
1. Compute raw Q*K scores
2. Concatenate sinks as a logit column
3. Softmax the combined logits
4. DROP the sink column (don't multiply by V)
5. Multiply by V
Our old code added sinks as a dummy zero-KV entry, which diluted
attention weights by adding an extra V=0 position to the softmax.
The HuggingFace reference (DeepseekV4ForCausalLM) applies an unweighted
RMSNorm after q_b_proj, normalizing Q before attention. Without it, Q
magnitudes are too large, causing attention scores to collapse to uniform
(entropy ~3.2 with 24 positions) and the model to produce garbage.
q_b_norm has no learnable parameters — just q / RMS(q).
This explains the nearly-uniform attention weights we've been seeing.
When enabled, bypasses mHC pre/post blocks and uses direct residual
connections with 0.1 scaling. This helps isolate whether the mHC
implementation is causing the garbage output.
The DSV4 Pro model uses rope_type='yarn' with factor=16. Our
build_rope_cache was using standard RoPE with theta=10000, completely
ignoring YaRN scaling. This produced wrong cos/sin values for all
positions, causing incorrect attention scores and garbage output.
YaRN modifies the RoPE frequencies:
- High-frequency components: unchanged
- Low-frequency components: scaled by 1/factor
- Medium: smooth interpolation
Config: factor=16, beta_fast=32, beta_slow=1, orig_max_pos=65536
