Gemma 4 E4B-it QAT, W4A16 with int4 embeddings, text only (unofficial)

The E4B text-only W4A16 repack xbill9/gemma-4-E4B-it-qat-q4_0-w4a16-ct-text with its embedding tables packed as int4 too, the E4B twin of xbill9/gemma-4-E2B-it-qat-q4_0-w4a16-ct-text-emb4. QAT put the embedding tables on the same 4-bit grid as the linears (group 32), so the packing recovers that grid; an off-grid group would abort the build.

Table	bf16	int4 + fp16 scales
`embed_tokens_per_layer` (per-layer embeddings)	5.250 GiB	1.477 GiB
`embed_tokens`	1.250 GiB	0.352 GiB
`lm_head`	tied to `embed_tokens`	0.352 GiB int4, untied copy
Checkpoint		4.26 GiB

0 off-grid groups in either table; 74.78 % (per-layer embeddings) and 73.48 % (embed_tokens) of values bit-identical, worst error 0.66 % of the group maximum.
lm_head is untied: vLLM ties the output layer by copying the embedding's .weight, which a packed embedding does not have, so the same levels and scales are written a second time as lm_head and run as an int4 linear. The model was trained tied.
Linear layers are unchanged from the W4A16 repack (0 group-level mismatches against the QAT source, verify_report.json).

Made with embed_int4.py (--embed-tokens, default fp16 scales). Needs vLLM 0.29 or later (CompressedTensorsEmbeddingWNA16Int).

Limitations